Machine learning-based methods for modeling pMHC conformers

Abstract

The present disclosure relates to methods for generating a plurality of compatible structures of a peptide-protein complex, which can be used, for example, to identify surface fingerprints and interfacial features of the peptide-protein complex. The methods can include inputting peptide sequence data and protein sequence data into a trained machine learning model to determine: (i) predicted pairwise distance data for peptide amino acid residues and protein amino acid residues in the peptide-protein complex; and (ii) predicted dihedral angle data for peptide amino acid residues in the peptide-protein complex; identifying an initial structure of the peptide-protein complex based on the predicted pairwise distance data and the predicted dihedral angle data; and generating the plurality of compatible structures of the peptide-protein complex based on the initial structure, the predicted pairwise distance data, and the predicted dihedral angle data.

Description

Cross-reference to related applications

[0001] This application claims priority to U.S. Provisional Patent Application Serial No. 63 / 567,360, filed March 19, 2024, the contents of which are incorporated herein by reference in their entirety.

[0002] Reference to electronic sequence listing The contents of the electronic sequence list (146392067640seqlist.xml; size: 2,756 bytes; and creation date: March 3, 2025) are incorporated herein by reference in their entirety. Technical Field

[0003] This disclosure generally relates to machine learning-based methods for modeling the conformational states of peptide-protein complexes, taking into account the dynamic properties of three-dimensional peptide and protein structures, and for predicting the characteristics of peptide-protein complexes. In some aspects, this disclosure more specifically relates to applying the disclosed methods to generate sets of conformational isomers of peptide-major histocompatibility complex (pMHC) complexes, which can be used to predict other characteristics of downstream peptide binding and / or peptide-MHC interactions. Background Technology

[0004] The binding interactions that drive peptide-protein complex formation are fundamental to a variety of biochemical processes essential to organisms. Examples include, but are not limited to, receptor-ligand binding-dependent cell signaling pathways and enzyme-substrate binding-dependent catalytic reactions. Peptide-protein binding interactions are also fundamental to the adaptive immune system in humans and other vertebrates. Human leukocyte antigen (HLA) is expressed as a cell surface receptor protein that binds antigenic peptides in a restricted manner and presents them to T cells, thereby allowing the differentiation of self-antigens from foreign antigens. The HLA complex is a gene complex located on chromosome 6 (Chr 6) that encodes cell surface proteins responsible for regulating the immune system. The major histocompatibility complex (MHC) is a linked set of genetic loci containing HLA genes that play fundamental roles, such as in cell-mediated immune responses to infectious diseases and in the reception of transplanted tissues. The MHC complex encodes the α-chains of MHC class I molecules HLA-A, HLA-B, and HLA-C (alleles), as well as the α-chains and β-chains of MHC class II molecules HLA-DR, HLA-DP, and HLA-DQ (allotypes), all of which are expressed in a codominant manner.

[0005] Cytotoxic T lymphocytes (cytotoxic T cells) are activated upon recognition of peptides that bind to MHC class I molecules (see, for example, Ito et al. 2010, “Regulation of the Induction and Function of Cytotoxic T Lymphocytes by Natural Killer T Cell”, Journal of Biomedicine and Biotechnology 2010:641757). Activated cytotoxic T cells secrete essential cell lysis mediators (e.g., perforin, granzymes, etc.) and induce apoptosis in target cells (tumor cells, virus-infected cells, etc.). Furthermore, activated cytotoxic T cells secrete cytokines such as interferon-γ (IFN-γ) and tumor necrosis factor-α (TNF-α), which enhance antigen presentation and mediate antipathogenic effects. Studies have shown that CD4 T cells producing various cytokines such as interleukin-2 (IL2) or IFN-γ are essential for generating effective cytotoxic T cell immunity against infections and cancer.

[0006] Activation of helper T (Th) cells upon recognition of peptide fragments (e.g., epitopes derived from protein antigens) that bind to MHC class II (MHC-II) molecules on antigen-presenting cells leads to the production of antigen-reactive antibodies. Sufficient affinity of the peptide for the MHC molecule is a prerequisite for immunogenicity (i.e., the ability of the peptide to trigger an immune response).

[0007] The ability to accurately predict the binding of peptides containing candidate MHC-I binding epitopes (typically 8–15 amino acid residues in length) and candidate MHC-II binding epitopes (typically 13–25 amino acid residues in length) to MHC-I / MHC-II molecules, as well as the immunogenicity of such peptide-MHC complexes (i.e., their ability to be recognized by T cell receptors and trigger an immune response), will be of significant value, for example, in identifying amino acid residues in proteins that can be safely altered to eliminate (or minimize) the immunogenicity of candidate therapeutic proteins and / or enhance the immunogenicity of candidate peptides for inclusion in vaccines, including candidate peptides derived from patient-specific neoantigens for the development of personalized cancer vaccines. Peptide-MHC binding affinity is primarily determined by the amino acid sequence of the peptide binding core (also known as the "minimum epitope," which refers to the portion of the peptide that extends between the "anchor residues" of the peptide, which are embedded in the pocket of the binding groove of the MHC molecule in the peptide-MHC complex, and typically have a median length of about nine amino acid residues); however, the N-terminal flanking (N-flanking) sequence and / or the C-terminal flanking (C-flanking) sequence (i.e., one or more residues on the N-terminal and / or C-terminal sides of the binding core) can also affect peptide-MHC binding.

[0008] Tools exist for predicting the binding of peptides to proteins to form complexes (e.g., tools for predicting the binding of peptides to MHC molecules to form pMHC complexes). However, existing tools often fail to accurately predict the properties of peptide-protein complexes (such as the immunogenicity of a specific pMHC). For example, peptides predicted to bind to MHC molecules often turn out to be false positives, meaning the resulting pMHC complex does not actually elicit a cellular immune response (i.e., the peptide predicted to bind to MHC molecules to form a pMHC complex is not actually immunogenic). Therefore, tools that can more accurately predict the properties of corresponding peptide-protein complexes remain needed. Summary of the Invention

[0009] While tools exist for predicting the ability of peptides to bind to proteins and form complexes, the inventors hypothesize that the limitations of these existing tools (i.e., their frequent inaccurate prediction of the properties of peptide-protein complexes, such as the immunogenicity of a particular pMHC) are at least partly due to their failure to account for the dynamic properties of peptide-protein complexes and the range of permissible conformational states. The inventors hypothesize that tools that consider the dynamic properties of three-dimensional peptide and protein structures would more accurately predict the properties of the corresponding peptide-protein complexes.

[0010] This paper discloses machine learning (ML)-based methods for modeling the conformational states of peptide-protein complexes, taking into account the dynamic properties of three-dimensional peptide and protein structures. The disclosed methods can be used to generate a set (i.e., a group) of permissible peptide-protein conformational isomers. In some examples, permissible peptide-protein conformational isomers are the allowed three-dimensional structures of the peptide-protein complex that can exist at any given time point and affect the function of the peptide-protein complex at any given time point. The set of permissible peptide-protein conformational isomers represents the permissible distribution of structural parameters of the peptide-protein complex and can be used to more accurately predict the characteristic and functional properties of the peptide-protein complex (e.g., complementary interface features, complementary surface fingerprint features, and / or binding of the peptide-protein complex to other proteins or protein complexes). The improved prediction accuracy achieved by the disclosed method and system is based on: using embedded representations of peptide and protein sequences as input; performing machine learning-based predictions of structural constraints on peptide-protein complexes (e.g., predicted posterior distributions of pairwise distance data for amino acid residues in the peptide and protein, and predicted posterior distributions of dihedral angle data for the peptide in the peptide-protein complex); and using predicted structural constraints to select an initial template structure for the peptide-protein complex and generate multiple compatible structures that capture the dynamic behavior of the peptide-protein complex in solution. As mentioned above, previous methods failed to account for the dynamic properties and range of permissible conformational states of peptide-protein complexes in solution. The inventors have demonstrated through extensive benchmarking on a ensemble of molecular dynamics (MD) simulations that this method is capable of recovering most of the backbone of peptide-HLA complexes involving peptide lengths up to 13 amino acids sampled by MD simulations. The inventors have further demonstrated that this method is capable of capturing the dynamic properties of wild-type and mutant peptide-HLA complexes recognized by TCR. To the inventors’ knowledge, the method described herein is the first to capture the flexibility of peptides (even those up to 13 amino acids long) within minutes without relying on time-consuming MD simulations.

[0011] In some aspects, for example, the disclosed machine learning-based methods can be used to generate a set of conformational isomers for pMHC complexes that represents the allowed distribution of structural parameters for a given pMHC complex and can be used downstream to predict the characterization and functional properties of pMHC complexes (e.g., complementary interface features, complementary surface fingerprint features, and / or predictions of the binding of pMHC complexes to other proteins or protein complexes). More accurate predictions of pMHC characterization and functional properties can, for example, facilitate the recognition of epitopes (e.g., peptide fragments derived from tumor-associated neoantigens) that elicit robust immune responses upon presentation on the surface of antigen-presenting cells. Improved representation of the dynamic properties of peptide-HLA complexes also improves the ability to capture conformations that may be present when peptide-MHC complexes are bound by TCRs. Dynamic allosteric changes are thought to play an important role in TCR recognition of peptide-HLA complexes, and correctly capturing this is important for accurately predicting TCR cross-reactivity and specificity. Therefore, the disclosed methods can help select peptide fragments for inclusion in or development of neoantigens for therapeutic cancer vaccines. In some instances, tumor-associated neoantigens can be patient-specific, and the disclosed methods can facilitate the selection of peptide fragments of patient-specific neoantigens for inclusion in or development of personalized cancer vaccines. The disclosed methods can also aid in the design and / or characterization of any TCR-based therapeutics or other tumor antigen recognition therapies. For example, the disclosed methods can facilitate the design and / or characterization of bispecific T-cell connectives (BiTEs), T-cell receptor-engineered T-cell (TCR-T) therapies, chimeric antigen receptor (CAR) T-cell therapies (CAR-T), and the like.

[0012] In other respects, this document also discloses systems configured to perform any of the methods disclosed herein, and computer-readable storage media comprising one or more programs including instructions that, when executed by one or more processors of the system, cause the system to perform any of the methods disclosed herein. Such systems may include one or more processors and one or more computer-readable storage media storing instructions that, when executed by the one or more processors, cause the one or more processors to perform any of the methods disclosed herein.

[0013] This paper discloses a computer-implemented method for generating multiple compatible structures of peptide-protein complexes. The method includes: inputting peptide sequence data of at least one peptide and protein sequence data of at least one protein into a trained machine learning model to determine: predicted pairwise distance data between at least one peptide amino acid residue of at least one peptide and at least one protein amino acid residue of at least one protein in the peptide-protein complex; and predicted dihedral angle data between at least one peptide amino acid residue of at least one peptide in the peptide-protein complex; identifying an initial structure of the peptide-protein complex based on the predicted pairwise distance data and the predicted dihedral angle data; and generating multiple compatible structures of the peptide-protein complex based on the initial structure, the predicted pairwise distance data, and the predicted dihedral angle data. The multiple compatible structures may also be referred to as a "set of conformational isomers." The initial structure may also be referred to as a "template structure." A trained machine learning model can be one that has been trained to take peptide sequence data of at least one peptide and protein sequence data of at least one protein as input, and to produce the following as output: predicted pairwise distance data between at least one peptide amino acid residue of at least one peptide and at least one protein amino acid residue of at least one protein in the peptide-protein complex; and predicted dihedral angle data for at least one peptide amino acid residue of at least one peptide in the peptide-protein complex. The machine learning model can be trained using structural data from multiple training peptide-protein complexes. For each of the multiple training peptide-protein complexes, the structural data can include the following values: pairwise distance between at least one peptide amino acid residue of the peptide and at least one protein amino acid residue of the protein in the training peptide-protein complex, and one or more dihedral angles of at least one peptide amino acid residue of the peptide in the training peptide-protein complex. These values ​​can be determined experimentally and / or predicted using a structural prediction algorithm.

[0014] In some embodiments, the computer-implemented method further includes predicting at least one auxiliary feature of the peptide-protein complex based on a plurality of compatible structures of the peptide-protein complex.

[0015] In some embodiments, each compatible structure of the peptide-protein complex conforms to one or more posterior distributions of predicted pairwise distance data and one or more posterior distributions of predicted dihedral data output by a trained machine learning model. For each of one or more pairs of at least one peptide amino acid residue of at least one peptide and at least one protein amino acid residue of at least one protein, the predicted pairwise distance data may include a predicted posterior distribution of the pairwise distances between the peptide amino acid residues and the protein amino acid residues in the peptide-protein complex. For each of one or more peptide amino acid residues in at least one peptide, the predicted dihedral data may include a predicted posterior distribution of one or more dihedral angles of the peptide amino acid residues in the peptide-protein complex. Predictive data including predicted posterior distributions may refer to predictive data including predicted values ​​of parameters characterizing the distributions. The posterior distributions may be used as constraints in peptide docking schemes, i.e., in a physics engine used to model peptide-protein complexes.

[0016] In some embodiments, a trained machine learning model is at least partially trained to predict at least one auxiliary feature of an input protein-peptide complex. In some embodiments, each compatible structure of the peptide-protein complex corresponds to sequence data of at least one peptide and at least one protein. In embodiments, generating multiple compatible structures of the peptide-protein complex based on an initial structure, predicted pairwise distance data, and predicted dihedral angle data includes generating a template structure that corresponds to the sequence data of at least one peptide and at least one protein using the initial structure and homology modeling methods.

[0017] In some embodiments, at least one auxiliary feature includes the prediction of the binding of at least one peptide to at least one protein. In some embodiments, at least one auxiliary feature includes the prediction of at least one interfacial feature of the peptide-protein complex.

[0018] In some embodiments, the peptide sequence data includes peptide embedding representations of at least one peptide, which are generated using a trained peptide embedding machine learning model. In some embodiments, the protein sequence data includes protein embedding representations of at least one protein, which are generated using a trained protein embedding machine learning model. A trained peptide embedding machine learning model can take a peptide sequence as input and produce a peptide embedding representation as output. A trained protein embedding machine learning model can take a protein sequence as input and produce a protein embedding representation as output. The trained peptide embedding machine learning model and / or the trained protein embedding machine learning model can be a machine learning model trained in a self-supervised manner to learn peptide / protein embeddings from input sequences. For example, the peptide embedding machine learning model and / or the trained protein embedding machine learning model can have been trained using unlabeled data (e.g., protein and / or peptide sequence data) and tasks such as masked language modeling, causal language modeling, etc.

[0019] In some embodiments, the trained peptide embedding machine learning model and the trained protein embedding machine learning model are the same model. In some embodiments, the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model includes a base model. In some embodiments, the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model includes a protein language model. In some embodiments, the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model includes a large language model (LLM). In some embodiments, the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model includes an evolutionary scale model (ESM). In some embodiments, the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model are integrated with a trained machine learning model and retrained simultaneously with the trained machine learning model. In embodiments, the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model are fine-tuned (e.g., after self-supervised training) for peptide-protein complex correlation prediction tasks. This training can be supervised training. In embodiments, the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model are fine-tuned for peptide-protein binding prediction. Fine-tuning of a task-specific model (e.g., a base model) may include incorporating the model into a machine learning model trained for the specific task and retraining the model's parameters at least partially for that task. The trained peptide embedding machine learning model and / or the trained protein embedding machine learning model may be a deep learning model, such as a deep neural network. Any deep learning architecture known in the art suitable for language modeling may be used.

[0020] In some embodiments, the trained machine learning model includes a trained neural network. In some embodiments, the trained machine learning model includes a trained multilayer perceptron (MLP), a trained attention-based neural network, or a trained state-space model.

[0021] In some embodiments, the trained machine learning model is trained using a training dataset that includes incomplete structural data of one or more peptide-protein complexes. In some embodiments, the trained machine learning model is trained using a training dataset that, for each of a plurality of training peptide-protein complexes, includes values ​​of pairwise distances between peptide amino acid residues and protein amino acid residues in the training peptide-protein complex, and values ​​of dihedral angles of peptide amino acid residues in the training peptide-protein complex, wherein the training data does not include all pairwise distances of one or more training peptide-protein complexes, and / or does not include dihedral angles ψi and ψj of one or more peptide amino acid residues in one or more training peptide-protein complexes. Both. In other words, this incomplete training peptide-protein complex structural data can still be used to train a machine learning model. In the example, the trained machine learning model has been trained using training data that includes the dihedral angle values ​​of the peptide amino acid residues in the training peptide-protein complex, but only includes the dihedral angles ψ and φ. The value of one of them.

[0022] In some embodiments, the trained machine learning model is further configured to determine the uncertainty in predicting pairwise distance data and predicting dihedral angle data. In embodiments, the trained machine learning model is configured to predict parameters characterized by: the distribution of pairwise distances between at least one peptide amino acid residue of at least one peptide in a peptide-protein complex and at least one protein amino acid residue of at least one protein; and the dihedral angle ψ of at least one peptide amino acid residue in at least one peptide in a peptide-protein complex. The joint distribution. In an embodiment, a trained machine learning model is configured to predict parameters characterized by: the distribution of pairwise distances between at least one peptide amino acid residue of at least one peptide and at least one protein amino acid residue of at least one protein in the peptide-protein complex; the distribution of the dihedral angle ψ of at least one peptide amino acid residue of at least one peptide in the peptide-protein complex; and the dihedral angle ψ of at least one peptide amino acid residue of at least one peptide in the peptide-protein complex. The distribution. Parameters characterizing the distribution may include parameters characterizing the maximum likelihood of the distribution (e.g., the mean) and parameters characterizing the uncertainty of the distribution (e.g., the dispersion) (e.g., the standard deviation or variance).

[0023] In some embodiments, the machine learning model is trained or has been trained by: receiving peptide sequence data of multiple peptides; receiving protein sequence data of multiple proteins; receiving peptide-protein binding data of multiple peptide-protein combinations; training a classifier based on the peptide sequence data, protein sequence data, and peptide-protein binding data to predict peptide-protein binding of specific combinations of peptides and proteins; and attaching a regressor head to at least a portion of the trained classifier and training the regressor head based on pairwise distance data and dihedral data derived from experimental (e.g., crystal) structural data of an exemplary peptide-protein complex set to predict pairwise distance data and dihedral data of the peptide-protein complex (e.g., probability distributions of pairwise distance data and dihedral data). In some embodiments, during the training of the regressor head, the pairwise distance data and dihedral data derived from experimental (e.g., crystal) structural data for each combination of a single peptide amino acid residue and a single protein amino acid residue in an exemplary peptide-protein complex are presented as separate training instances. In some embodiments, the machine learning model has been trained in at least two steps. In the first step, the machine learning model may have been trained to predict peptide-protein bindings using input peptide sequence data and input protein sequence data. This training data includes: peptide sequence data for multiple peptides; protein sequence data for multiple proteins; and peptide-protein binding data for multiple peptide-protein combinations. Predicting peptide-protein bindings may include classifying peptide-protein combinations as binding / not binding to form peptide-protein complexes. This training data may include peptide and protein sequence data for multiple peptides and proteins known to bind to each other to form peptide-protein complexes, as well as peptide and protein sequence data for multiple peptides and proteins whose binding to each other is unknown or not expected to form peptide-protein complexes. In the first step, the machine learning model may include a peptide sequence embedding model, a protein sequence embedding model, and a classification head. This model may be referred to as a "classifier." In the second step, at least a portion of the machine learning model (e.g., the peptide sequence embedding model and the protein sequence embedding model) may have been trained to predict pairwise distance data and dihedral data (e.g., the probability distributions of pairwise distance data and dihedral data for peptide-protein complexes). The training in the second step can be based on pairwise distance data and dihedral angle data derived from experimental (e.g., crystal) structural data of an exemplary (i.e., training) set of peptide-protein complexes. In the first step, the machine learning model can include a peptide sequence embedding model, a protein sequence embedding model, and a regression head. This model can be referred to as a "regressor".

[0024] In some embodiments, the machine learning model (e.g., a regressor) is further trained or has already been further trained based on enhanced pairwise distance data and dihedral data. Enhanced pairwise distance data and dihedral data can refer to data that has not yet been experimentally determined. Such data can be obtained, for example, using structural modeling methods, such as by using a trained structure prediction machine learning model, such as models from the AlphaFold family. This data can be obtained by using a trained structure prediction machine learning model to obtain predicted crystal structures of peptide sequences and protein sequences known to bind together to form peptide-protein complexes. In some embodiments, the enhanced pairwise distance data and dihedral angle data are derived by: receiving peptide sequence data and protein sequence data of at least one peptide and at least one protein known to bind and form a peptide-protein complex; processing the peptide sequence data and protein sequence data of at least one peptide and at least one protein using a previous version of a trained machine learning model to determine the uncertainty of the pairwise distance data and the uncertainty of the dihedral angle data of the peptide-protein complex; applying the determined uncertainty of the pairwise distance data and the determined uncertainty of the dihedral angle data to the structure of an exemplary peptide-protein complex to generate multiple compatible structures of the peptide-protein complex; and extracting the pairwise distance data and dihedral angle data of the multiple compatible peptide-protein complex structures to be used as enhanced training data for a regressor.

[0025] In some embodiments, identifying an initial structure of a peptide-protein complex based on predicted pairwise distance data and predicted dihedral angle data includes: obtaining multiple experimental (e.g., crystal) structures of the peptide-protein complex from a protein structure database; and selecting an experimental (e.g., crystal) structure from the multiple experimental (e.g., crystal) structures based on sequence and / or structural similarity between the experimental structure and the peptide-protein complex. Structural similarity can be the similarity between pairwise distance data and dihedral angle data determined for the crystal structure and predicted pairwise distance data and predicted dihedral angle data determined by a trained model. In embodiments, the experimental structure is selected based on the similarity between the pairwise distance data and dihedral angle data determined for the crystal structure and the maximum likelihood estimates of the predicted pairwise distance data and predicted dihedral angle data determined by a trained model. In embodiments, selecting experimental structures based on structural similarity includes: (i) for each exemplary peptide-protein experimental structure among a plurality of experimental structures, determining the likelihood (e.g., mean likelihood, mean log-likelihood) of the distances and dihedrals in the peptide-protein experimental structure using the predicted distributions of pairwise distances and dihedrals of the peptide-protein complex output by a trained machine learning model; and (ii) selecting the exemplary peptide-protein experimental structure with the highest likelihood. In some embodiments, the protein structure database includes a protein database (PDB). In embodiments, the initial structure includes a plurality of initial structures.

[0026] In some embodiments, generating multiple compatible structures of a peptide-protein complex based on an initial structure, predicted pairwise distance data, and predicted dihedral angle data includes: inputting the initial structure of the peptide-protein complex into a structure modeling software package; inputting the predicted pairwise distance data and predicted dihedral angle data determined by a trained machine learning model for the peptide-protein complex into the structure modeling software package; and iteratively sampling and evaluating a structure scoring function from multiple possible structures of the peptide-protein complex to identify multiple compatible structures of the peptide-protein complex, wherein a compatible structure of the peptide-protein complex is a structure that: (i) is compatible with one or more constraints imposed by the predicted pairwise distance data and predicted dihedral angle data, and (ii) optimizes the structure scoring function. In embodiments, one or more constraints imposed by the predicted pairwise distance data and predicted dihedral angle data are included in the structure scoring function.

[0027] In some embodiments, iterative sampling from multiple possible structures of the peptide-protein complex is performed using a Markov Chain Monte Carlo (MCMC) sampling algorithm. In some embodiments, the MCMC sampling algorithm includes the Metropolis-Hastins algorithm. In some embodiments, the MCMC sampling algorithm includes the Hamiltonian Monte Carlo algorithm. In some embodiments, the MCMC sampling algorithm includes the Langevin Monte Carlo algorithm.

[0028] In some embodiments, the structure scoring function includes a proxy for free energy calculation. In some embodiments, the structure scoring function includes a proxy for calculating the free energy of peptide-protein binding, a proxy for calculating the free energy of at least one peptide-protein interface feature, or any combination thereof. In some embodiments, the scoring function is based on one or more of the following: ab initio quantum mechanical calculations, density functional theory (DFT) calculations, semi-empirical calculations, molecular mechanical force field calculations, statistical potential calculations, neural potential calculations, or machine learning models trained on structural data. In some embodiments, the structure modeling software package includes FlexPepDock.

[0029] In some embodiments, the computer-implemented method further includes iteratively traversing a plurality of candidate peptides to identify at least one peptide having the highest likelihood for forming a peptide-protein complex.

[0030] In some embodiments, at least one peptide includes at least a portion of a peptide expressed in normal cells of a subject. In some embodiments, at least one peptide includes at least a portion of a neoantigen expressed in tumor cells from a cancer patient. In some embodiments, at least one peptide includes at least a portion of a peptide expressed in a virus or bacteria.

[0031] In some embodiments, at least one protein comprises a major histocompatibility complex (MHC) class I protein. In some embodiments, at least one protein comprises a major histocompatibility complex (MHC) class II protein. In some embodiments, the computer-implemented method further comprises, for at least one MHC class I or II protein, iteratively traversing a plurality of candidate peptide sequences or portions thereof expressed in normal cells from a subject, to order the peptide sequences or portions thereof according to the likelihood of formation of a peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complex. In some embodiments, the computer-implemented method further comprises, for at least one MHC class I or II protein, iteratively traversing a plurality of candidate neoantigen sequences or portions thereof expressed in tumor cells from a cancer patient, to order the neoantigen sequences or portions thereof according to the likelihood of formation of a peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complex. In some embodiments, the computer-implemented method further includes iteratively traversing a plurality of candidate peptide sequences or portions thereof expressed in viruses or bacteria for at least one MHC class I or II protein to order the peptide sequences or portions thereof according to the likelihood of formation of a peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complex.

[0032] In some embodiments, the computer-implemented method (or methods related to the computer-implemented method) further includes developing a personalized anticancer vaccine based on the sequencing of a novel antigen sequence or a portion thereof. In some embodiments, the computer-implemented method (or methods related to the computer-implemented method) further includes developing an antiviral vaccine, an antibacterial vaccine, or an autoreactive vaccine based on the sequencing of a peptide sequence or a portion thereof. In some embodiments, the computer-implemented method (or methods related to the computer-implemented method) further includes predicting one or more characteristics of a complex comprising a peptide-protein complex and additional proteins. The additional proteins may be, for example, T cell receptors, antigen recognition molecules, T cell adaptor molecules, etc. The predicted characteristics may include one or more of the following: the likelihood of complex formation, the stability of the complex, the binding energy of the complex, and the total energy of the complex.

[0033] This document also discloses systems comprising: one or more processors; and a memory communicatively coupled to the one or more processors and configured to store instructions that, when executed by the one or more processors, cause the system to perform any of the computer-implemented methods described herein.

[0034] This document also discloses a non-transitory computer-readable storage medium that stores one or more programs, the programs including instructions that, when executed by one or more processors of the system, cause the system to perform any of the computer-implemented methods described herein.

[0035] The terms and expressions used are descriptive rather than restrictive, and their use is not intended to exclude any equivalents of the features shown and described or portions thereof, but it should be recognized that various modifications are possible within the scope of the claimed invention. Therefore, it should be understood that although the claimed invention has been disclosed through specific exemplary embodiments and optional features, modifications and variations of the concepts disclosed herein can be adopted by those skilled in the art, and such modifications and variations are considered to be within the scope of the invention as defined by the appended claims.

[0036] Incorporated by reference All publications, patents, and patent applications mentioned in this specification are incorporated herein by reference in their entirety, to the extent that each individual publication, patent, or patent application is specifically and individually indicated as being incorporated herein by reference in its entirety. In the event of any conflict between terminology used herein and terminology found in the incorporated references, the terminology used herein shall prevail. Attached Figure Description

[0037] This disclosure is described in conjunction with the accompanying drawings: Figure 1 A block diagram of an exemplary prediction system based on some implementations of the disclosed methods and systems is provided.

[0038] Figure 2 Non-limiting examples of block diagrams of systems for generating a set of conformational isomers of peptide-protein complexes according to some embodiments of the disclosed methods and systems are provided.

[0039] Figure 3A Schematic diagrams are provided illustrating how peptide sequences are converted into peptide embedding representations using a protein language model, based on some implementations of the disclosed methods and systems.

[0040] Figure 3B Schematic diagrams are provided illustrating how protein sequences (e.g., MHC sequences) are converted into protein embedding representations using protein language models, based on some implementations of the disclosed methods and systems.

[0041] Figure 4Schematic diagrams are provided of some implementations of the disclosed methods and systems, using a trained machine learning model (e.g., a structural constraint prediction model) to output pairwise distance data and dihedral data of peptide-protein complexes (e.g., pMHC complexes) based on input embedding representations of peptide and protein (e.g., MHC) sequences.

[0042] Figure 5 A schematic diagram is provided showing how, according to some embodiments of the disclosed methods and systems, a set of conformational isomers of peptide-protein complexes (e.g., pMHC complexes) is generated using pairwise distance data and dihedral data of peptide-protein complexes (e.g., peptide-MHC complexes) output by a trained machine learning model (e.g., a structural constraint prediction model).

[0043] Figure 6 A non-limiting schematic diagram is provided of one embodiment of the disclosed method, illustrating the process of training a machine learning model to predict pairwise distance data and dihedral angles of peptide-protein complexes.

[0044] Figure 7 A non-limiting schematic diagram is provided of one embodiment of the disclosed method, illustrating a process of using a trained machine learning model to predict pairwise distance data and dihedral angles of peptide-protein complexes, which can then be used to: (i) select template structures from a protein structure database, and (ii) provide constraints on permissible conformations when generating a set of conformational isomers of peptide-protein complexes. MHC pseudosequence YFAMYGEKV…(SEQ ID NO:1); peptide sequence LLFGYPVYV (SEQ ID NO:2).

[0045] Figure 8 Non-limiting examples are provided, demonstrating how the predicted distributions of pairwise distance data and dihedral angle data of pMHC complexes can be converted into structural constraints, and how a set of pMHC conformations can then be generated based on such structural constraints.

[0046] Figure 9 A non-limiting example of a process flow diagram for generating multiple compatible structures of peptide-protein complexes according to one embodiment of the disclosed method is provided.

[0047] Figure 10 A non-limiting example of a process flowchart for training a machine learning model to predict the distribution of pairwise distance data and dihedral angle data of peptide-protein complexes is provided according to one embodiment of the disclosed method.

[0048] Figure 11ANon-limiting examples of data on the root mean square deviation (RMSD) of predicted dihedral angles Φ and Ψ in a benchmark pMHC complex set containing peptides of different lengths are provided, wherein the dihedral angles Φ and Ψ are processed separately during training and inference.

[0049] Figure 11B Non-limiting examples of comparative data on the root mean square deviation (RMSD) of predicted dihedral angles Φ and Ψ in a benchmark pMHC complex set containing peptides of different lengths are provided, wherein the dihedral angles Φ and Ψ are processed individually or jointly during training and inference.

[0050] Figure 11C Non-limiting examples of comparative data for predicting the root mean square deviation (RMSD) of the dihedral angles Φ and Ψ in a benchmark pMHC complex set containing peptides of different lengths are provided, one dataset (red symbol) was generated using a previous protein structure prediction model (AlphaFold – FineTune), and the other dataset (green symbol) was generated using one implementation of the method disclosed herein.

[0051] Figure 11D Non-limiting examples are provided for predicting the mean dihedral distance in a benchmark pMHC complex set containing peptides of different lengths.

[0052] Figure 11E A non-limiting example is provided of preliminary data for recovering the crystal structure contacts of a benchmark pMHC complex set using the disclosed structure prediction method.

[0053] Figure 12A Non-limiting examples are provided for recovering distance data of pMHC complex test groups using the disclosed structure prediction method.

[0054] Figure 12B Non-limiting examples are provided for recovering dihedral angle data of a pMHC complex test group using the disclosed structure prediction method.

[0055] Figure 12C Non-limiting examples are provided for recovering dihedral data of a representative pMHC 9-mer example (1I1Y) using the disclosed structural prediction method.

[0056] Figure 13 Non-limiting examples are provided of the number of models that overlap with corresponding baseline molecular dynamics simulations in the set of conformational isomers predicted using the methods of this disclosure or baseline methods that do not use ML to predict structural constraints.

[0057] Figure 14A Non-limiting examples of evaluation metrics are provided for comparing sets of conformational isomers obtained using the disclosed structure prediction method or baseline methods without ML predictions with corresponding benchmark molecular dynamics simulations in the first test set.

[0058] Figure 14B A non-limiting example of an evaluation metric is provided for comparing a set of conformational isomers obtained using the disclosed structure prediction method or a baseline method without ML predictions to corresponding benchmark molecular dynamics simulations in a second test set.

[0059] Figure 15A Non-limiting examples are provided of the ability of the disclosed structure prediction method to capture the dynamic behavior of peptides involved in TCR recognition.

[0060] Figure 15B Data for the same peptide as in Figure 15B are provided, demonstrating that molecular dynamics (MD) simulations cannot overcome the energy barrier associated with side chain flipping in the TCR-binding conformation of the peptide.

[0061] Figure 16A A non-limiting example of reference data for pMHC template structure selection is provided.

[0062] Figure 16B A non-limiting example of reference data for pMHC template structure refinement is provided.

[0063] Figure 17A Non-limiting examples of molecular dynamics data for peptide conformation sampling within pMHC complexes are provided.

[0064] Figure 17B Provided Figure 17A Non-limiting example of a graphical representation of a pMHC complex.

[0065] Figure 17C Provided Figure 17A A non-limiting example of a graph showing the variation of peptide backbone fluctuations with residue numbering in a pMHC complex.

[0066] Figure 18A Non-limiting examples of molecular dynamics data for peptide conformation sampling within pMHC complexes are provided.

[0067] Figure 18B Provided Figure 18A Non-limiting example of a graphical representation of a pMHC complex.

[0068] Figure 18C Provided Figure 18AA non-limiting example of a graph showing the variation of peptide backbone fluctuations with residue numbering in a pMHC complex.

[0069] Figure 19 Non-limiting examples of block diagrams of computer systems according to some implementations of the methods and systems disclosed herein are provided.

[0070] Figure 20 Some implementations of the methods and systems disclosed herein are provided as... Figure 4 This is a non-limiting example of a block diagram of an artificial intelligence (AI) architecture included as part of an exemplary computing system.

[0071] In the accompanying drawings, similar parts and / or features may have the same reference numerals. Furthermore, various parts of the same type can be distinguished by adding a dash after the reference numerals and a second reference numeral to differentiate similar parts. If only the first reference numeral is used in the description, the description applies to any similar parts having the same first reference numeral, regardless of the second reference numeral. Detailed Implementation

[0072] Machine learning-based methods for modeling the conformational states of peptide-protein complexes are described, taking into account the dynamic properties of three-dimensional peptide and protein structures. The disclosed methods can be used to generate a set of permissible peptide-protein conformational isomers (e.g., the set of permissible three-dimensional structures that a peptide-protein complex may have at any given time point – also referred to herein as the “compatible structure” of the peptide-protein complex, which can affect the function of the peptide-protein complex at any given time point). The set of permissible peptide-protein conformational isomers can be represented as a permissible distribution of the structural parameters of the peptide-protein complex. The set of permissible peptide-protein conformational isomers can be used to more accurately predict the characteristic and functional properties of peptide-protein complexes (e.g., complementary interface features, complementary surface fingerprint features, and / or binding of the peptide-protein complex to other proteins or protein complexes).

[0073] In some instances, for example, the disclosed method for generating multiple compatible structures of peptide-protein complexes may include a first step of: inputting peptide sequence data (e.g., amino acid sequence or embedded representation thereof) of at least one peptide and protein sequence data (e.g., amino acid sequence or pseudo-sequence, or embedded representation thereof) of at least one protein (which may be a full-length protein or a truncated form thereof, including at least the portion of the protein that interacts with the peptide) into a trained machine learning model (e.g., a structure-constrained prediction model, or a machine learning model comprising a structure-constrained prediction model and one or more embedding models configured to obtain protein embedding representations and peptide embedding representations, also referred to herein as “protein embeddings” and “peptide embeddings”). The structural constraint prediction model can be configured (e.g., trained) to process embedding representations of peptide and protein sequences to determine: (i) predicted pairwise distance data (or its distribution) for at least one peptide amino acid residue of at least one peptide in a peptide-protein complex and at least one protein amino acid residue of at least one protein (i.e., one or more predicted distances for each of a pair or more pairs of residues containing peptide and protein residues, or parameters of the predicted distribution of said distances, wherein the distance between residues can be represented as the distance between predetermined atoms of the residues (such as Cα of the residue); and (ii) predicted dihedral angle data (or its distribution) for at least one peptide amino acid residue of at least one peptide in a peptide-protein complex (i.e., one or more predicted values ​​and / or dihedral angles ψ for each of one or more peptide residues in a peptide-protein complex). One or more predicted values, or parameters of the corresponding distribution of the dihedral angles). The predicted structural constraint data can then be used to identify the initial structure of the peptide-protein complex and to generate multiple compatible structures of the peptide-protein complex based on the initial structure and the predicted values ​​or distributions of pairwise distances and dihedral angles.

[0074] In some instances, the method may further include predicting the binding of a peptide-protein complex to another protein or protein complex based on multiple compatible structures of the peptide-protein complex. For example, in some instances, predicting the binding of a peptide-protein complex based on multiple compatible structures of the peptide-protein complex includes: identifying complementary interface features between the peptide-protein complex and another protein or protein complex; identifying complementary surface fingerprint features between the peptide-protein complex and another protein or protein complex; identifying potential representations of the peptide-protein complex and another protein or protein complex; and providing the complementary interface features, complementary surface fingerprint features, and potential representations of the peptide-protein complex and another protein or protein complex as input to a second trained machine learning model, which is configured (e.g., trained) to use the input and output to predict the binding of the peptide-protein complex to another protein or protein complex.

[0075] The improvements in processing performance and prediction accuracy achieved by the disclosed methods and systems are based on: using embedded representations of peptide and protein sequences as input; performing machine learning-based predictions of structural constraints on peptide-protein complexes (e.g., predicted pairwise distance data or posterior distributions of amino acid residue pairwise distance data in peptides and proteins, and predicted dihedral angle data or posterior distributions of dihedral angle data of peptides in peptide-protein complexes); and using predicted structural constraints to select an initial template structure for the peptide-protein complex and generate multiple compatible structures that capture the dynamic behavior of the peptide-protein complex in solution. As mentioned above, previous methods failed to account for the dynamic properties of peptide-protein complexes in solution and the range of permissible conformational states.

[0076] ML-based generation for conformational isomer sets of peptide-protein complexes Figure 1Block diagrams of exemplary prediction systems according to some embodiments are provided. Prediction system 100 may be used, for example, to predict the structural properties of peptide-protein complexes (e.g., pMHC complexes), determine multiple permissible conformations (e.g., a set of conformational isomers) of peptide-protein complexes (e.g., pMHC complexes), and / or determine the interaction of pMHC with one or more other proteins in an immunoprotein complex (IPC) related to the immunological activity of the peptide (such as, for example, the interaction between pMHC and a T cell receptor molecule (TCR) or a portion thereof), and particularly mutant peptides (e.g., neoantigen peptides). Prediction system 100 includes a computing platform 102, a data storage 104, and a display system 106. Computing platform 102 may take various forms. In some embodiments, computing platform 102 includes a single computer (or computer system) or multiple computers communicating with each other. In some embodiments, computing platform 102 may be a cloud computing platform.

[0077] Data storage 104 and display system 106 each communicate with computing platform 102. In some examples, one or more of data storage 104 or display system 106 may be considered part of computing platform 102 or otherwise integrated with it. Therefore, in some examples, computing platform 102, data storage 104, and display system 106 may be separate components that communicate with each other; however, in other instances, combinations of these components may be integrated together. Communication between the different components can be achieved using any number of wired communication links, wireless communication links, optical communication links, or combinations thereof.

[0078] The prediction system 100 includes a sequence analyzer 108, which can be implemented using hardware, software, firmware, or a combination thereof. In some embodiments, the sequence analyzer 108 can be implemented in a computing platform 102. The sequence analyzer 108 receives sequence data 110 for processing. For example, the sequence data 110 can be sent as input to the sequence analyzer 108, retrieved from data storage 104 or some other type of storage (e.g., cloud storage), accessed from cloud storage, or obtained in some other way. In some cases, it can be in response to receiving input from a user via an input device ( Figure 1 (Not shown) User input is used to retrieve sequence data 110 from data storage 104.

[0079] Sequence data 110 can be generated from the processing of sample set 112. Sample set 112 can take the form of one or more biological samples from one or more subjects (e.g., diseased samples, healthy samples, or combinations thereof). Sample set 112 can include samples obtained from a subject's tumor (e.g., biopsy or tumor resection samples), or samples containing tumor genetic material from the subject (e.g., liquid biopsy). The tumor can manifest as, for example, lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, gastric cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, non-small cell lung cancer, small cell lung cancer, or combinations thereof.

[0080] The samples in sample set 112 may include, for example, various IPC molecules, various peptides, nucleic acids encoding said IPC molecules and / or peptides, or combinations thereof. When sample set 112 includes diseased samples, peptides may include one or more mutant peptides (e.g., neoantigens). IPC molecules may include, for example, various MHC molecules, various TCR molecules, or combinations thereof.

[0081] In some embodiments, sample set 112 includes IPC 114 (e.g., MHC class I molecules, MHC class II molecules, various TCR molecules, etc.) and / or nucleic acid sequences encoding IPC 114 from which the sequence of IPC 114 can be inferred. Furthermore, the sample set may include at least one protein 123 (i.e., the source protein) or nucleic acid sequences encoding protein 123 or a portion thereof from which the sequence of protein 123 (referred herein to as protein sequence 160) or a portion thereof (such as amino acid chain 116 or peptide 118) can be inferred. Amino acid chain 116 can be identified from the at least one protein 123, and the amino acid chain may be an amino acid chain including peptide 118 (the smallest epitope). Amino acid chain 116 may optionally additionally include N-flanking 120 and / or C-flanking 122, which refer to sequences located on the N-terminal side and C-terminal side of peptide 118, respectively. Peptide 118 can be considered a mutant peptide when it includes one or more variants (e.g., one or more sequence variations) compared to its corresponding reference sequence. Protein 123 can be the source protein of amino acid chain 116, which can be generated through proteolysis, the process by which a protein (e.g., protein 123) is broken down into smaller polypeptides or amino acids. Protein 123 can be cleaved into smaller polypeptides or amino acids by specific enzymes called proteases, which cleave the peptide bonds between amino acids in protein 123.

[0082] Sample set 112 can be processed to generate sequence data 110. In some embodiments, multiple samples in sample set 112 can be processed at different times. In some embodiments, prediction system 100 includes a sample analyzer that can be used to process sample set 112 to generate sequence data 110. Sequence data 110 includes, for example, at least one amino acid sequence 129 and at least one IPC sequence 124 (e.g., an IPC sequence 124 corresponding to IPC 114). Amino acid sequence 129 may include one or more of the following: peptide sequence 126 (e.g., a peptide sequence 126 corresponding to peptide 118), N-terminal flanking (N-flanking) sequence 128 (e.g., an N-flanking sequence 128 corresponding to N-flanking 120), or carboxyl-terminal flanking (C-flanking) sequence 130 (e.g., a C-flanking sequence 130 corresponding to C-flanking 122). One or more subsequences of amino acid sequence 129 (e.g., peptide sequence 126, N-flanking sequence 128, and C-flanking sequence 130) can be treated individually or as a single sequence.

[0083] When IPC 114 is MHC, IPC sequence 124 may be, for example, an MHC sequence 135 characterizing at least a portion of MHC. When IPC 114 is TCR, IPC sequence 124 may be, for example, a TCR sequence 131 characterizing at least a portion of TCR. In some embodiments, IPC sequence 124 may include an MHC sequence 135 characterizing at least a portion of an MHC molecule and a TCR sequence 131 characterizing at least a portion of a TCR molecule. In some embodiments, sequence data 110 may include IPC sequence 124 in the form of MHC sequence 135 (which characterizes at least a portion of an MHC molecule) and a separate TCR sequence 131 (which characterizes at least a portion of TCR).

[0084] Protein sequence 160 characterizes at least a portion of protein 123. In some embodiments, protein sequence 160 can be identified by a reverse lookup in a database (e.g., the UniProt database) based on mutant peptide data (e.g., IPC sequence 124) obtained from a sample. Protein sequence 160 can be used to provide contextual and / or orthogonal information when characterizing peptide 118, such as comparing peptide 118 with its wild-type counterpart to determine the expression of protein 123 in one or more samples or tissues. The availability of protein sequence 160 is entirely optional, and the methods of this disclosure may use only, for example, peptide sequence 118, peptide sequence 116, or portions thereof.

[0085] Peptide sequence 126 characterizes at least a portion of peptide 118. N-flanking sequence 128 characterizes at least a portion of N-flanking sequence 120. In some embodiments, when the number of amino acids (or amino acid residues) upstream of the N-terminus of the peptide is large, the corresponding sequence of N-flanking sequence 120 may be pruned to generate N-flanking sequence 128. In other words, N-flanking sequence 128 may comprise or consist of a predetermined number of residues upstream of peptide 118 (i.e., the N-terminus). C-flanking sequence 130 characterizes at least a portion of C-flanking sequence 122. In some embodiments, when the number of amino acids (or amino acid residues) downstream of the C-terminus is large, the corresponding sequence of C-flanking sequence 122 may be pruned to generate C-flanking sequence 130. In other words, C-flanking sequence 130 may comprise or consist of a predetermined number of residues downstream of peptide 118 (i.e., the C-terminus).

[0086] Sequence analyzer 108 receives sequence data 110 as input for processing. Sequence analyzer 108 includes one or more machine learning models 132 that process the sequence data 110. In some embodiments, sequence analyzer 108 may process the sequence data 110 (or its embedded form) before sending it to one or more of the machine learning models 132 for further processing (e.g., using embedding engine 140). In some embodiments, machine learning model 132 may include one or more of, for example, one or more embedding engines 140, a structural constraint prediction model 141, and a conformational isomer set generation engine 142.

[0087] Embedding generation engine 140 can be configured to process peptide and / or protein sequence (e.g., MHC sequence) data (e.g., IPC sequence 124 and / or amino acid sequence 129) and output embeddings of the corresponding sequences. As used herein, the embedding of a peptide or protein sequence can refer to a vector representation of the peptide or protein sequence that encodes important features of the peptide or protein in a manner that facilitates processing by downstream computational models. The embedding of a peptide or protein sequence can be a latent representation of a machine learning model trained to learn a latent space in which an input sequence can be represented and from which a sequence can be reconstructed and / or one or more features of the input sequence can be predicted. Structure constraint prediction model 141 can be configured to process the peptide and / or protein embeddings of the peptide-protein complex output by embedding generation engine 140 and output a set of predicted structure constraints associated with the peptide-protein complex. The set of predicted structure constraints may include, for example, predicted pairwise distance data (or their distribution) between at least one peptide amino acid residue and at least one protein amino acid residue in the peptide-protein complex and / or predicted dihedral data (or their distribution) between at least one peptide amino acid residue in the peptide-protein complex. The conformational isomer set generation engine 142 can be configured to generate multiple compatible structures (i.e., conformational isomer sets 148) of the peptide-protein complex based on an initial template structure and a set of structural constraints output by the structure constraint prediction model 141. Therefore, the conformational isomer set generation engine 142 can be configured to receive an initial template structure and a set of structural constraints for the peptide-protein complex output by the structure constraint prediction model 141 as input, and generate multiple compatible structures of the peptide-protein complex as output. These multiple compatible structures can form a conformational isomer set, also referred to as a conformation set, structure set, or simple set. The conformational isomer set of the peptide-protein complex can be a set of conformations that collectively describe the structure of the peptide-protein complex. The conformational isomer set can be expected to capture a range of conformations that the peptide-protein complex can adopt under normal (e.g., physiological) conditions.

[0088] One or more machine learning models 132 can be used in either training or prediction mode. In training mode, one or more of the machine learning models 132 can be trained using training data 133. Examples of training data 133 are described in more detail below. One or more machine learning models 132 are trained so that they can be used in prediction mode.

[0089] One or more of the machine learning models 132 are used to process the IPC sequence 124 and the amino acid sequence 129. In some embodiments, a separate processing engine (such as, for example, a separate embedding engine 140) may be used to process the IPC sequence and the amino acid sequence. This can achieve improved predictive performance of one or more machine learning models 132. In some embodiments, one or more of the machine learning models 132 process one or more of the following: MHC sequence 135, TCR sequence 131, protein sequence 160, peptide sequence 126, N-flanking sequence 128, or C-flanking sequence 130. Examples of implementations for these different processing engines are described in more detail below.

[0090] As used herein, the terms “processing engine,” “engine,” and “model” identify at least one software component and / or a combination of at least one software component and at least one hardware component designed / programmed / configured to interact with and / or transmit data to other software and / or hardware components, including but not limited to other processing engines.

[0091] One or more machine learning models 132 process sequence data 110 to generate output that can be used to generate report 144. Report 144 may include the exact output of any one or more of the machine learning models 132, a transformed (e.g., further processed) or filtered version of the output of any one or more of the machine learning models 132, or both. In some cases, report 144 may include notifications, suggestions, alerts, or other information generated by sequence analyzer 108 based on the output of one or more of the machine learning models 132.

[0092] Report 144 may be the following output: it includes, for example, information about the structural constraints of the peptide-protein complex (e.g., the pMHC complex), such as information output by or derived from the structural constraint prediction model 141; a set 148 of conformational isomers of the peptide-protein complex (e.g., the pMHC complex), such as a set 148 of conformational isomers output by the set generation model 142; and predicted features of the peptide-protein complex (e.g., the pMHC complex), such as predicted features derived, for example, at least in part, from the set 148 of conformational isomers output by the set generation model 142, which may include, for example, the immunological activities of interest associated with one or more peptides (e.g., one or more mutant peptides). For example, report 144 may include information 146 regarding structural predictions of peptide-protein complexes (e.g., pMHC complexes) (e.g., predicted distributions of pairwise distances and / or dihedral angles of amino acid residues), information 148 regarding the set of conformational isomers of peptide-protein complexes (e.g., the set of allowed conformations consistent with structural predictions), peptide-protein complex feature predictions 162, immunological activity predictions related to amino acid chains 116 (e.g., peptide 118, N-flanking 120, C-flanking 122, etc.) and IPC 114 (e.g., MHC-I, MHC-II, TCR, etc.), or any combination thereof. Report 144 may include, for example, interaction information (e.g., interaction affinity predictions predicting the binding affinity between the peptide and MHC, or interaction predictions predicting whether an MHC allele or allotype will present the peptide on the cell surface), immunogenicity information (e.g., immunogenicity predictions predicting the ability of the peptide to elicit an immune response in the context of an MHC molecule), or both. Interaction information can provide predictions about a selected set of interactions between amino acid sequence 129 and IPC sequence 124. Immunogenicity information can provide predictions about the immunogenicity of amino acid sequence 129 (e.g., including the immunogenicity of peptide 118). Immunogenicity information can provide predictions about the immunogenicity of peptide 118 in the context of IPC 114 (where IPC is an MHC molecule). Immunogenicity information can provide predictions about the immunogenicity of peptide 118 (or a peptide-protein complex containing peptide 118 and IPC 114, where IPC is an MHC molecule) relative to IPC 114, where IPC is a TCR molecule.

[0093] In some embodiments, report 144 may be displayed on a graphical user interface (GUI) 150 on display system 106. A user can view and / or interact with report 144 via GUI 150. In some embodiments, a user can use report 144 to make treatment decisions regarding a subject who is a subject from at least one of the samples obtained (or collected) from sample set 112. In some embodiments, a user can use report 144 to make decisions regarding the manufacture of treatment for a subject who is a subject from at least one of the samples obtained (or collected) from sample set 112. For example, a user can use report 144 to select peptide sequences for manufacturing vaccines (such as, for example, personalized cancer vaccines).

[0094] In some embodiments, the prediction system 100 sends report 144 to a remote system 152 (e.g., wirelessly). The remote system 152 may be a cloud computing platform, cloud storage, another computer system, user equipment (e.g., a smartphone, tablet, laptop, etc.), or some other type of platform. In some embodiments, the remote system 152 may be a therapeutic manufacturing system (or machine) or a part thereof.

[0095] Figure 2 A non-limiting example of a block diagram of a system 200 (e.g., a computer-implemented system) for generating a set of conformational isomers of peptide-protein complexes is provided. System 200 can be used... Figure 1 The prediction system 100 described herein shall be used for implementation. For example, system 200 may include... Figure 1 One or more of the sequence analyzer 108 and the machine learning model 132.

[0096] The inputs to system 200 include protein sequence data of at least one protein (e.g., MHC sequence 135) and peptide sequence data of at least one peptide (e.g., peptide sequence 126).

[0097] In some instances, the input may include peptide sequence data of at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, or more than 50 peptides or portions thereof. In some instances, the peptide sequence data may include sequence data of peptides with a length of at least 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 amino acid residues. In some instances, the peptide may include normal peptides (e.g., peptides expressed in normal cells) or portions thereof. In some instances, the peptide may include mutant peptides (e.g., neoantigens expressed in tumor cells) or portions thereof. Normal peptides and / or mutant peptides can be isolated from a subject (e.g., a patient) or otherwise identified from samples previously obtained from the subject (such as, for example, by analyzing nucleic acid sequence data obtained from a sample of the subject containing normal and / or tumor-derived nucleic acids). In some instances, peptides may include bacterial peptides or portions thereof expressed by bacteria. In some instances, peptides may include viral peptides or portions thereof expressed in viruses.

[0098] In some instances, the input may include protein sequence data of at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 100, 1000, or more than 1000 proteins or portions thereof. In some instances, the protein sequence data may include sequences or pseudo-sequences of at least 20, 22, 24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52, 54, 56, 58, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, or at least 200 amino acid residues. In some instances, the protein may include an IPC protein. In some instances, an IPC protein (also referred to herein as an IPC molecule) may include, for example, an MHC-I molecule, an MHC-II molecule, a TCR molecule, or a combination thereof.

[0099] The input protein sequence (e.g., MHC sequence 135) and peptide sequence 126 can be processed by the embedding generation engine 140 to generate protein embedding representations (e.g., MHC embedding representation 206) and peptide embedding representations 208. The protein embedding representations (such as MHC embedding representation 206) and peptide embedding representation 208 can be represented as data structures in vector form. For example, each of the embedding representations can be represented as a vector of size (t, e), where the protein (e.g., MHC sequence 135) is represented by a vector of size (t, e). Figure 1Each residue of IPC 114 and peptide 118 shown is represented by a token t, and the contextual information of each residue is encoded in the (e) dimension. In some instances, in addition to the "residue tokens," the embedding generation engine 140 can add "special tokens," which can be used to indicate the start or end of a protein sequence. For an embedding representation instantiated as a vector of size (t, e), t can represent the total number of residue tokens and special tokens. The protein embedding representation encodes important features of a protein or peptide in a manner that can be processed by a downstream computational model.

[0100] A non-exhaustive list of features that can be implicitly encoded in protein embedding representations (i.e., each of these features has been shown to be a property that can be captured from the protein embedding representation obtained from the protein / peptide embedding model and can be captured against the input protein sequence) includes: Amino acid sequence: The amino acid sequence that makes up a protein. This sequence determines the protein's structure and function.

[0101] Structural properties: various aspects of the three-dimensional structure of a protein, such as α-helices, β-sheets, and other secondary structures.

[0102] Functional properties: Information about the biological function of a protein, such as its enzyme activity, binding site, or role in cellular processes.

[0103] Evolutionary information: Proteins with similar sequences or structures are often evolutionarily related. Some features can capture evolutionary relationships and similarities, which can be useful for tasks such as protein classification or functional prediction.

[0104] Interactions with other molecules: This can also encode information about how proteins interact with other molecules, such as DNA, RNA, or small molecules, such as drugs. This includes binding sites and interacting domains.

[0105] Post-translational modifications: Proteins often undergo post-translational modifications (such as phosphorylation or glycosylation), which affect their function. These modifications can sometimes be inferred from the sequence and context, and therefore can be encoded as characteristics.

[0106] Physical and chemical properties: Attributes such as the hydrophobicity, charge, size, and shape of amino acid residues can affect how proteins fold and interact with other molecules.

[0107] Contextual information: In some models, the context in which a protein or part of a protein appears can be encoded (such as in a specific type of cell or a specific metabolic pathway).

[0108] In some instances, the embedding generation engine 140 includes a protein embedding model 202 for receiving an MHC sequence 135 and generating an MHC embedding representation 206, and a peptide embedding model 204 for receiving a peptide sequence 126 and generating a peptide embedding representation 208. In some instances, peptide embedding model 204 and protein embedding model 202 are the same model; in some instances, peptide embedding model 204 and protein embedding model 202 are different models. Exemplary methods for generating peptide and / or protein sequence embeddings are described, for example, in Ofer et al. (2021), “The Language of Proteins: NLP, Machine Learning & Protein Sequences”, Computational and Structural Biotechnology Journal 19:1750-1758.

[0109] In some instances, peptide embedding model 204 and / or protein embedding model 202 may include large language models (LLMs) or protein language models (PLMs), such as deep learning algorithms configured to process strings to perform predictive or generative tasks (such as recognition, summarization, translation, prediction, and content generation). LLMs and PLMs (which are special cases of LLMs trained using protein sequence data) are typically trained, at least partially, in an unsupervised manner (i.e., trained using unlabeled data and tasks such as masked language modeling, causal language modeling, etc.) on very large datasets (e.g., very large datasets of peptide and / or protein sequences). LLMs and PLMs are typically (but not limited to) transformer-based deep neural networks. Large language models are described in more detail in, for example, the following: Li et al. (2021), “BioSeq-BLM: APlatform for Analyzing DNA, RNA and Protein Sequences Based on Biological Language Models”, Nucleic Acids Res. 49, e129–e129, and Naveed et al. (2023), “A Comprehensive Overview of Large Language Models”, arxiv.org / pdf / 2307.06435.pdf.

[0110] In some instances, peptide embedding model 204 and / or protein embedding model 202 may include evolutionary scale models (ESMs), such as transformer-based LLMs, which use masked language modeling trained on over 250 million unlabeled data of different protein sequences (specifically, using all sequence data from UniParc - see UniProt Consortium, The universal protein resource (UniProt). Nucleic AcidsRes. 36, D190–D195 (2008)). ESMs have been shown to generate vector representations for input protein sequences that can encode the biochemical properties of amino acids in the sequence, biological variation and long-range homology of the sequence, as well as the secondary structure and tertiary contacts of the sequence (see, for example, Rives et al. (2021), “Biological Structure and Function Emerge from Scaling Unsupervised Learning to 250 Million Protein Sequences”, PNAS118(15)e2016239118).

[0111] Then, MHC embedding representation 206 and peptide embedding representation 208 are used as input to a trained machine learning model (structural constraint prediction model 141), which is configured to process the embedding sequence data and output, for example, pairwise distance data 210 (i.e., predicted pairwise distance data between at least one peptide amino acid residue of the peptide and at least one protein amino acid residue of the protein in the peptide-protein complex) and dihedral data 212 (i.e., predicted dihedral data between at least one peptide amino acid residue of the peptide in the peptide-protein complex).

[0112] In some instances, predicting pairwise distance data and predicting dihedral angle data may include a predicted posterior distribution for each parameter. The posterior distribution may be characterized by, for example, the mean, standard deviation, or any other measure of uncertainty / variability (or any other set of statistical parameters used to characterize the distribution, depending on the type of distribution being inferred). In some instances, the structurally constrained prediction model 141 may be further configured to determine the mean, standard deviation, and / or uncertainty (or any other set of statistical parameters used to characterize the distribution) of the predicted pairwise distance data and / or predicted dihedral angle data.

[0113] In some instances, the structural constraint prediction model 141 may include a trained neural network, as described in more detail elsewhere herein. In embodiments, the trained neural network is a deep neural network. In some instances, the trained neural network may include, for example, a trained multilayer perceptron (MLP), a trained attention-based neural network, a trained recurrent neural network, a trained state-space model, or other suitable types of neural networks or deep learning models. In embodiments, the structural constraint prediction model 141 includes a multilayer perceptron.

[0114] The pairwise distance data 210 and dihedral data 212 (along with the initial template structure of the pMHC complex) can then be used as input to an ensemble generation engine 142, which is configured to use the pairwise distance data 210 and dihedral data 212 as constraints and, starting from the initial template structure, generate multiple compatible structures of the pMHC complex (i.e., a set of conformational isomers 148), wherein the compatible structures of the pMHC complex are those that are (i) compatible with one or more constraints imposed by the predicted pairwise distance data 210 and predicted dihedral data 212, and (ii) have an optimized structure scoring function, as will be discussed in more detail below.

[0115] In some instances, the method may further include iteratively traversing multiple candidate peptide sequences or portions thereof for at least one MHC class I or II protein. In some embodiments, the candidate peptide sequences or portions thereof may be ordered based on the likelihood of forming a peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complex, and / or using one or more features of one or more peptide-protein complexes (such as, for example, a peptide-MHC-TCR complex, or other complexes containing peptide-MHC and an antigen recognition molecule) predicted by a set 148 of conformational isomers obtained for the respective complexes. For example, a metric such as a score estimated from one or more energy functions calculated for the set of conformational isomers may be used to determine whether a peptide-MHC complex is likely to form, to quantify the stability of the peptide-MHC complex (e.g., by quantifying the energy of the peptide-MHC complex using an energy function), and / or to quantify the binding of the peptide to the MHC molecule (e.g., by quantifying the energy of the peptide-MHC interface). Furthermore, such measures can be used to compare the differences among multiple peptides in terms of their likelihood of forming a peptide-MHC complex with a given set of one or more MHC molecules, the stability of their complex with said MHC molecules, and / or their binding affinity with said MHC molecules. Therefore, scores estimated from one or more energy functions calculated for multiple sets of conformational isomers (each set containing a given MHC allele and candidate peptides with candidate peptide sequences) can be used to rank candidate peptide sequences, such as to obtain, for example, a ranking indicating one or more of the following: the likelihood of these candidate peptides forming a complex with a given MHC allele, the stability of the complex containing the candidate peptide and the given MHC allele, and the binding affinity of the candidate peptide with the given MHC allele. Any energy function known in the art can be used, including but not limited to the Rosetta energy function described, for example, by Cornell et al. (J Am Chem Soc. 1996;118(9):2309–2309). In some instances, the peptide may include a normal peptide (e.g., a peptide expressed in normal cells) or a portion thereof isolated from a subject (e.g., a patient). In some instances, the peptide may include a mutant peptide (e.g., a neoantigen expressed in tumor cells) or a portion thereof isolated from a subject (e.g., a patient). In some instances, the peptide may include a bacterial peptide or a portion thereof expressed by bacteria. In some instances, the peptide may include a viral peptide or a portion thereof expressed in a virus.

[0116] In some instances, the method may further include developing personalized anticancer vaccines, antiviral vaccines, antibacterial vaccines, or autoreactive vaccines based on the sequencing of peptide sequences or portions thereof. For example, the method may include selecting one or more candidate peptides for inclusion in a vaccine based on sequencing.

[0117] Figure 3A An example process for generating peptide sequence embeddings, according to some embodiments, is described. References Figure 3A Protein Language Model (PLM) 302 (corresponding to Figure 2 The peptide embedding model 204 can receive peptide sequences and output peptide sequence embeddings 208 (corresponding to...). Figure 2 The peptide embedding in the image is represented as 208. A protein language model (PLM) is a machine learning model that can be based on natural language processing methods (e.g., deep learning models), such as attention-based models (including, but not limited to, transformers), and can be trained on a set of unlabeled protein sequences. A protein language model can be trained to understand (i.e., using self-supervised learning, sometimes called “pre-training”), and can also optionally be trained to predict one or more properties of peptides and / or proteins based on the amino acid sequences that form such peptides and / or proteins (sometimes called fine-tuning, task-specific training, or downstream task training). In some instances, a protein language model can infer a range of features from an amino acid sequence, including the primary, secondary, tertiary, and quaternary structures of the peptide and / or protein (where applicable). In the case of proteins, PLMs can be used, for example, to predict how proteins fold, what domains are present, where active sites are located, and how stable the protein is. In some cases, PLMs can also predict protein-protein, protein-peptide, and protein-nucleic acid interactions, post-translational modifications, and the effects of mutations. In some cases, PLMs can identify intracellular localization signals, understand evolutionary relationships, and predict protein function. Similarly, PLMs can provide insights into the dynamic behavior of proteins and identify potential drug binding sites, which are valuable for drug discovery and understanding the molecular basis of diseases. In some embodiments, PLM 302 may include evolutionary scale modeling (ESM model) or a variant of an ESM model. In some embodiments, PLM 302 may include ProteinBERT, UniRep, or other suitable types of PLMs (e.g., any protein language model known in the art).

[0118] In some embodiments, PLM 302 includes a pre-trained protein language model, such as a pre-trained ESM model. A pre-trained PLM may refer to a PLM that has been trained in a self-supervised manner to learn representations of protein sequences (i.e., to learn a potential space from which protein sequences can be reconstructed, e.g., reconstructing ground truth sequences from damaged sequences by training the model) and has not yet been further trained or fine-tuned for a specific prediction task. In some embodiments, the input peptide sequence 126 may include a sequence of amino acid residues, and PLM 302 may be configured to obtain multiple embeddings (i.e., vector representations) by obtaining corresponding embeddings for each amino acid residue. The model may be further configured to obtain a single embedding 208 by aggregating the multiple embeddings corresponding to the amino acid residue sequence (e.g., by performing element-wise averaging).

[0119] Figure 3B Exemplary processes for generating MHC sequence (or other protein sequence) embeddings, according to some embodiments, are described. References Figure 3B PLM 304 (which can be related to Figure 3A The PLM 302 shown in the image is either the same model or a different model, and corresponds to... Figure 2 The protein embedding model 202 can receive MHC sequence 135 and output MHC sequence embedding 206 (corresponding to...). Figure 2The MHC embeddings in PLM 304 are represented as 206. As discussed herein, PLMs can be trained on a large number of proteins and thus can encode useful information and context about the input sequence represented by the sequence embeddings. In some embodiments, PLM 304 may include an evolutionary scale modeling (ESM model) or a variant of an ESM model. In some embodiments, PLM 304 may include ProteinBERT, UniRep, etc. In some embodiments, PLM 304 includes a pre-trained protein language model, such as a pre-trained ESM model. In embodiments, both PLM 304 and PLM 302 include an ESM model, a ProteinBERT model, a UniRep model, or any variant of these. In embodiments, both PLM 304 and PLM 302 include the same protein language model. This ensures that the embeddings of peptides and proteins come from the same model, thereby enhancing the ability to learn associations between peptides and proteins. In some embodiments, the MHC sequence 135 may include multiple amino acids that make up the corresponding allele. The multiple residues may be continuous or discontinuous. MHC sequence 135 can be a pseudo-sequence. For each MHC allele (also referred to herein as an “MHC molecule” or “HLA allele” or “HLA molecule”), the pseudo-sequence can be a 34-amino acid sequence typically located within 4 Å of the peptide that binds to the MHC allele. Pseudo-sequences of many MHC molecules are known in the art and described, for example, in Nielsen et al., PLoSOne. 29 Aug 2007; 2(8):e796. PLM 304 can be configured to obtain multiple embeddings (i.e., vector representations) by obtaining corresponding embeddings for each amino acid. The model can be further configured to obtain a single embedding 206 by aggregating multiple embeddings corresponding to amino acid residue sequences (e.g., by performing element-wise averaging).

[0120] Figure 4 A schematic diagram is provided showing the pairwise distance data and dihedral data of peptide-protein complexes (e.g., pMHC complexes) based on input peptide and protein (e.g., MHC) embeddings using a trained machine learning model (e.g., structural constraint prediction model 141).

[0121] Peptide embedding representation 208 and MHC embedding representation 206 are provided as input to a structure-constrained prediction model 141, which can be configured to infer the posterior distribution of pairwise distance data 210 of peptide and protein amino acid residues in peptide-protein (e.g., pMHC) complexes (e.g., the predicted pairwise C-values ​​of at least one peptide amino acid residue and at least one protein amino acid residue in the peptide-protein complex). α – C α The distribution of distance data), and dihedral data 212 of peptide amino acid residues in the pMHC complex (e.g., the distribution of predicted dihedral data for at least one peptide amino acid residue in a peptide-protein complex). In an embodiment, the structure-constrained prediction model 141 is configured (i.e., trained) to predict one or more parameters characterizing the distribution (e.g., posterior distribution) of the corresponding distribution of distances between corresponding pairs of peptide and protein amino acid residues in the peptide-protein complex, such as, for example, pairwise C-values ​​between at least one peptide amino acid residue and at least one protein amino acid residue in the peptide-protein complex. α – C α At least one distribution of distance. In an embodiment, the structural constraint prediction model 141 is configured (i.e., trained) to predict one or more parameters of the corresponding distribution (e.g., posterior distribution) of one or more dihedral angles characterizing each of one or more peptide amino acid residues in a peptide-protein complex. In an embodiment, the structural constraint prediction model 141 is configured (i.e., trained) to predict two dihedral angles (e.g., distances) characterizing each of one or more peptide amino acid residues in a peptide-protein complex. One or more parameters of the corresponding joint distribution (e.g., posterior distribution) of the peptide-protein complex and its corresponding amino acid residues (ψ). In an embodiment, the structural constraint prediction model 141 is configured (i.e., trained) to predict both: (i) one or more parameters characterizing the corresponding distribution of the distances between corresponding pairs of peptide and protein amino acid residues in the peptide-protein complex, and (ii) one or more parameters characterizing the corresponding distribution (e.g., posterior distribution) of one or more dihedral angles of each of the one or more peptide amino acid residues in the peptide-protein complex. In an embodiment, instead of predicting distributions, the structural constraint prediction model 141 may be trained to predict individual values ​​(e.g., samples of the distributions of pairwise distances and / or dihedral angles) that can fit the corresponding distributions to the individual values ​​to identify the distributions. The predicted pairwise distance data and predicted dihedral angle data may include a predicted posterior distribution for each parameter, which may be characterized, for example, by the mean, standard deviation, or any other set of statistical parameters used to characterize the distribution. In other words, the predicted pairwise distance data and / or predicted dihedral angle data may include one or more parameters characterizing one of one or more corresponding distributions for each pairwise distance and / or dihedral angle for which a prediction is made. The reference to predicted pairwise distances between amino acid residues can refer to the distances between selected heavy atoms in the residue backbone. The selected heavy atom can be the α-carbon (C6) of the corresponding residue. α Therefore, "paired distance" can be the distance between the α-carbons of two amino acid residues. These two amino acid residues include amino acid residues in peptides and amino acid residues in proteins. References to predicted dihedral angles can refer to the distance between peptide residues (e.g., peptide amino acid residues in a pair of residues, which contains both peptide amino acid residues and protein amino acid residues). Prediction of the dihedral angle ψ and / or ψ.

[0122] In some instances, the structural constraint prediction model 141 may include a trained neural network, as described in more detail elsewhere herein. In some instances, the trained machine learning model may include, for example, a trained multilayer perceptron (MLP), a trained attention-based neural network, or a trained graph neural network.

[0123] Figure 5 A schematic diagram is provided showing how a set 148 of conformational isomers of peptide-protein complexes (e.g., pMHC complexes) is generated using pairwise distance data 210 and dihedral data 212 of peptide-protein complexes (e.g., peptide-MHC complexes) output by a trained machine learning model (e.g., structural constraint prediction model 141).

[0124] Using predicted pairwise distance data 210 and dihedral angle data 212, an initial template structure 502 for a pMHC complex is selected from an exemplary set of pMHC experimental structures (such as crystal structures, e.g., crystal structures available in a protein structure database, such as a protein database (PDB)) based on the maximum likelihood estimate (MLE) of the predicted posterior distributions of pairwise distances and dihedral angles. In an embodiment, the initial template structure 502 for the peptide-protein complex is selected from an exemplary set of peptide-protein experimental structures, each of which contains peptides of the same length (i.e., the same number of amino acids) as peptides in a peptide-protein (e.g., peptide-MHC) complex. In an embodiment, selecting the initial template structure 502 from the exemplary set of peptide-protein experimental structures includes: (i) for each exemplary peptide-protein experimental structure in the set, determining the average likelihood (which can actually be calculated as a log-likelihood) of the distances and dihedral angles in the peptide-protein experimental structure using the predicted distributions of the pairwise distances and dihedral angles of the peptide-protein complex; and (ii) selecting the exemplary peptide-protein experimental structure with the highest average likelihood. This leads to the selection of the template closest to the peptide-protein complex based on the predicted structural constraints. The average likelihood of the distance and dihedral angle in the experimental peptide-protein structure can be calculated as follows: , where x is the distance or dihedral angle in the experimental structure, and The predicted parameters for the distance and dihedral angle distribution, as predicted by the model, are shown here as Gaussian parameters. In an embodiment, the initial template structure 502 of the peptide-protein complex is selected from an exemplary set of peptide-protein experimental structures, each of which is a representative experimental structure of a cluster of peptide-protein experimental structures. In an embodiment, the initial template structure 502 of the peptide-protein complex includes multiple template structures. In some such embodiments, each template structure is selected from an exemplary set of peptide-protein experimental structures, and each peptide-protein experimental structure is a representative experimental structure of a cluster of peptide-protein experimental structures. For example, multiple clusters of peptide-protein experimental structures can be obtained using clustering algorithms and predetermined structural similarity measures, each cluster containing peptides of the same length as the peptides in the peptide-protein complex to be modeled. Representative experimental structures can be identified for each cluster, thereby obtaining multiple representative experimental structures. One or more, or each, of the representative experimental structures can be used as the initial template structure 502. The representative experimental structure can be the cluster centroid, such as the structure closest to the corresponding cluster center using, for example, a predetermined distance measure. These clusters can be obtained through structural clustering of peptide backbones in an exemplary peptide-protein experimental structure set, which contains peptides of a predetermined length (e.g., the length of the peptide for which an initial template structure is being sought). These clusters can be obtained using agglomerative clustering methods and a predetermined distance metric. The predetermined distance metric can be the cosine difference of the dihedral angles of the peptides. This can be calculated, for example, for two peptides A and B. Where L is the length of the peptide. The initial template structure 502 can then be used as input to the integrated generation engine 142. The term "experimental structure" refers to the experimentally determined atomic 3D coordinates of a peptide-protein complex. Such structures can be determined using, for example, X-ray crystallography or nuclear magnetic resonance (NMR). As used herein, the term "crystal structure" refers to the experimentally determined atomic 3D coordinates of a peptide-protein complex via X-ray crystallography.

[0125] The predicted pairwise distance data 210 and dihedral angle data 212 also provide structural constraints on the permissible conformations of the pMHC complex and are also provided as input to the set generation engine 142.

[0126] Then, the ensemble generation engine 142 can be used to generate multiple compatible structures (e.g., a set of conformational isomers 148) of the pMHC complex based on the initial template structure 502 and constraints provided by the predicted pairwise distance data 210 and dihedral angle data 212. A compatible structure of the pMHC complex is one that is (i) compatible with one or more constraints imposed by the predicted pairwise distance and dihedral angle data, and (ii) optimizes the structure scoring function. Generating multiple compatible structures may include, for example, inputting the initial pMHC template structure into a structure modeling software package (e.g., FlexPepDock (Rosetta, www.rosettacommons.org), as described below: “Sub-angstrom modeling of complexes between flexible peptides and globular proteins”, Proteins, Vol. 78, No. 9, July 2010, pp. 2029-2040), and further inputting the predicted pairwise distance and dihedral angle data of the pMHC complex into the structure modeling software package. Structural modeling packages can be configured to iteratively sample from multiple possible structures of the pMHC complex (e.g., using a Markov chain Monte Carlo (MCMC) sampling algorithm) and evaluate each sampled structure using a structural scoring function (e.g., a surrogate for free energy calculations, as discussed in more detail elsewhere in this paper) to identify multiple compatible structures. Such a process is known, for example, described in Raveh et al., Proteins, Vol. 78, No. 9, July 2010, pp. 2029–2040. Predicted pairwise distance data and dihedral angle data can be used as constraints in peptide docking schemes (also known as “refinement schemes” or “refinement procedures”). Predicted pairwise distance data and dihedral angle data can be used for energy functions used in peptide docking schemes. For example, predicted pairwise distance data can be used as harmonic constraints in peptide docking schemes (e.g., it can be introduced as a harmonic function into the energy function used in peptide docking schemes – see, for example, Alford et al., J. Chem. Theory Comput. 2017, 13, 6, 3031–3048). Predicted dihedral angle data can be used as cyclic harmonic constraints in peptide docking schemes (e.g., it can be introduced as a cyclic harmonic function into the energy function used in peptide docking schemes – see, for example, Alford et al., J. Chem. Theory Comput. 2017, 13, 6, 3031–3048). In embodiments, generating multiple compatible structures of peptide-protein complexes involves pre-packaging the template to remove internal collisions in the protein and peptide.In embodiments, generating multiple compatible structures of peptide-protein complexes includes optimizing the peptide backbone structure relative to the receptor protein (e.g., using the MCMC algorithm with a minimization method), along with on-the-fly sidechain optimization. In embodiments, the peptide backbone optimization is repeated a predetermined number (N) times, thereby generating a number (N) of independent optimization trajectories. In embodiments, the peptide docking scheme uses coarse-grained models of the peptide and protein. In embodiments, the peptide backbone and peptide sidechains are iteratively optimized (e.g., using the MCMC algorithm), in conjunction with the rigid body placement of the peptide relative to the protein. In embodiments, the initial structural model (e.g., the coarse-grained model) uses a template structure (i.e., an experimentally derived template peptide-protein complex structure) for homology modeling. In embodiments, the initial structural model (e.g., the coarse-grained model) uses a method including: aligning the peptide sequence with the peptide sequence in the template structure, and using a threaded scheme (such as, for example, implemented in Rosetta) to generate a homology model using the alignment and template structure coordinates. In this embodiment, the homology model undergoes a restricted refinement phase, where only residues of the peptide and residues in the protein that are within a predetermined distance (e.g., 3.5 Å) of the peptide residues are refined (e.g., in the corresponding energy function in a Rosetta force field or any other structure modeling environment). In this embodiment, the initial structure model is then used to generate multiple compatible structures using the structure modeling software package described above.

[0127] In some instances, the set of conformational isomers 148 can be used for downstream processing to, for example, predict the characteristics of pMHC complexes (e.g., interfacial features, surface complementarity features, etc.).

[0128] Training a classifier to predict peptide-MHC binding Figure 6 A non-limiting schematic diagram is provided illustrating the process of training a machine learning model to predict pairwise distance data and dihedral angles of pMHC complexes. For example... Figure 6 As shown in the non-limiting example, the model is configured to accept MHC embedding representation 608 and peptide embedding representation 606 as inputs, the MHC embedding representation for a fixed-length MHC pseudo-sequence 604 with 34 residues, and the peptide embedding representation for a peptide sequence 602 with a variable length. The task performed by the trained model can be a probabilistic regression of the structural properties of the pMHC complex. In particular, the model can be trained to estimate the posterior distribution of pairwise distances between peptides and MHC amino acid residues, as well as the dihedral angles of the peptide residues. For example, if the parameterization of the posterior distribution consists of a decomposed Gaussian distribution, and the peptide is, for example, 9-mer, the model outputs the Gaussian mean and standard deviation parameters of the pairwise distances between MHC and peptide residues (a total of 34x9 distances), and eight... Similar parameters to the ψ dihedral angle.

[0129] For this example, the model architecture consists of a large pre-trained model (such as an evolutionary scale model (ESM)) (see, for example, Rives et al. (2021), “Biological Structure and Function Emerge from Scaling Unsupervised Learning to 250 Million Protein Sequences”, PNAS 118(15)e2016239118; Verkul et al. (2022), “Language Models Generalize Beyond Natural Proteins”, bioRxiv 2022-12), and includes additional layers specifically for probabilistic regression tasks.

[0130] The training is conducted in two phases, such as Figure 6 As shown in the figure. In the first stage, the ESM part of the model was fine-tuned for peptide-MHC binding classification, 610. A multilayer perceptron (MLP) with a single (logit) output was attached to the tandem ESM embedding of peptides and MHC. Specifically, the embedding of the ESM part of the model was averaged at residue positions and processed by an MLP head with a single (logit) output (labeled "classifier"). The inclusion of the MLP stage was to take advantage of the large number of publicly available pMHC sequences annotated with binary binding tags. For example, the model can be pre-trained on a labeled dataset of 2 million sequence pairs compiled by Chu et al. (2022), "A Transformer-Based Model to Predict Peptide–HLA Class I Binding and Optimize Mutated Peptides for Vaccine Design", Nature Machine Intelligence 4(3):300-311. It can be expected that the classification task provides an understanding of efficient binding interfaces of various peptides and MHC for models pre-trained on general proteins. Binary cross-entropy / log loss function It provides a measure of the difference between the predicted binary result and the actual binary label, used for training. Here, N is the number of data points, and y i Let p(y) be the actual label in the training data point distribution (q(y)), and p(y) be the actual label in the distribution (q(y)). i ) represents the data point yi The probability of belonging to the positive class predicted by the model. It quantifies the difference between the two distributions and promotes model training by penalizing inaccurate predictions.

[0131] In the second phase of training to predict pairwise distances and dihedrals, the classification multilayer perceptron (MLP) is discarded, and a new MLP, labeled "Regressor" 612, is introduced. Embedding models 606 and 608 can be frozen at this stage. While not mandatory, this can be advantageous because the structural data used to train the model in the second phase is typically less than that used in the first phase, thus reducing the likelihood of the model losing the generalization power learned in the first phase. Supervised training for the regression task requires crystal structure data of the pMHC complex (e.g., atomic coordinates of atoms in each amino acid residue of the protein), and only a few hundred such examples are available in the Protein Database (PDB). Each training instance is defined by data for a single MHC residue and a single peptide residue. Given a residue pair, the model extracts slices of the corresponding residues from the ESM embeddings, concatenates them, and passes the resulting tensor through the Regressor MLP. In other words, the model's forward calls extract slices corresponding to a given residue pair from the ESM embeddings (where "slices" are either peptide residue or MHC residue embeddings), concatenate them, and pass the resulting tensor through the new MLP. Each residue pair of the peptide and pseudo-MHC sequence is generated. The model output includes pairwise distance data 614 and dihedral data 616, and varies according to the parameterization used to characterize the predicted posterior distribution. The model output may include the predicted distribution of Cα distances for all amino acid residue pairs (e.g., the number of distributions equals the number of residues in the MHC molecule pseudo-sequence (typically 34) multiplied by the peptide length). The model output may include multiple residues in the peptide corresponding to the peptide length minus 1. The joint distribution of the dihedral angle and the ψ-dihedral angle (since the dihedral angle lies between the first residue at the N-terminus and the subsequent residue at the C-terminus of the first residue). For example, if the parameterization used consists of a joint Gaussian, then for a given residue pair, the model can output three means and a 3 x 3 covariance matrix controlling the ternary Gaussian, where the diagonal elements of the 3 x 3 covariance matrix include the squares of the three standard deviations, and the off-diagonal elements are the covariance values. If the parameterization used consists of a decomposed (diagonal) Gaussian, the output can be simplified to, for example, a set of three means and standard deviations. Note that it is not limited to using a Gaussian model to characterize the posterior distribution of the predicted pairwise distance and dihedral angle data; other options include Gaussian mixture models (joint or decomposed), von Mises distributions for angles, and their mixtures. The loss function used to train the new MLP 612 is negative log-likelihood (NLL): Evaluation is performed at distance and angle labels, where N is also the number of data points, and p(y i ) represents the data point y i The probability of having a true label (i.e., the joint probability of ground truth distance and dihedral angle in the predicted posterior distribution). The values ​​of mean and standard deviation can then be used as constraints in the peptide-protein docking scheme explained elsewhere in this paper. In this example, the model is trained to output the predicted distribution of pairwise distances between peptides and HLA residues, as well as the φ and ψ dihedral angles of the peptide residues. This can be expressed as: Where L is the peptide length and N is the protein sequence (e.g., pseudo-sequence) length. It is a pairwise distance matrix, and The terms are peptide dihedrals, and all terms are conditioned on the training set. In this example, the prediction distribution is parameterized as a Gaussian mixture with K components (although other distributions can be used as explained above). This can be expressed as: Where N(.|μ,σ) represents a Gaussian density with mean and variance parameters μ and σ, such that It is a parameter that controls the Gaussian component k of each distance element, and also Control each angle element. Coefficient For mixed weights, in k The sum of the above is 1, and for Similarly, the value of K (the number of components) can be set independently for each prediction distribution, or it can be set to the same value for multiple (e.g., all) prediction distributions. The value of K can be set as a hyperparameter, for example, using cross-validation. In the embodiment, the model is trained to predict a set of distribution parameters for input protein sequence embeddings and input peptide sequence embeddings, the distribution parameters characterizing the distribution of pairwise distances between peptide amino acid residues and protein amino acid residue pairs. The distance and the joint distribution of dihedral angles of peptide amino acid residues. Using the above notation, and in an embodiment where the predicted distributions of distance and dihedral angles are parameterized as a Gaussian mixture with K components, this can be represented as In an embodiment, the structural constraint prediction model (e.g., the new MLP 612) is trained using maximum likelihood estimation. In another embodiment, the structural constraint prediction model (e.g., the new MLP 612) is trained using a loss function defined as follows: Where w is a hyperparameter, L is the peptide length, N is the protein sequence (e.g., pseudo-sequence) length, and the subscript θ refers to the parameters of the structure-constrained prediction model (e.g., the new MLP 612). The hyperparameter w can be set to a predetermined value. In an embodiment, the structure-constrained prediction model (e.g., the new MLP 612) is trained using a loss function that weights the training loss corresponding to a first predetermined length of peptide with more training data to be less than the training loss corresponding to a second predetermined length of peptide with less training data.

[0132] In some instances, such as when crystal structure data for one or more peptide-protein complexes (e.g., one or more pMHC complexes) is incomplete (e.g., dihedral angles and / or pairwise distances of one or more amino acid residues are missing), a masking step can be implemented to exclude the missing information from the training data used to train and / or optimize the model. Furthermore, for ψ angles that are available but... When the angle is unavailable, amino acid residues in The joint distribution on ψ can be decomposed as follows: Similarly, for the angle ψ, it is possible to use but When the angle is unavailable, amino acid residues in The joint distribution on ψ can be decomposed as follows: The distribution θ can be optimized using either the left-hand or right-hand expression. Evaluating the left-hand expression requires... The measured values ​​of both and ψ. In the expression on the right (decomposition), evaluate the first term (conditional distribution). or It also needs The measurements of both ψ and ψ are used, but the evaluation focuses on the second term (marginal distribution). or Then each only needs Or the measured value of ψ. This allows it to still be based on only having Or, the model is trained using the measured values ​​of amino acid residues, ψ. The above decomposition is straightforward for Gaussian distributions, but for other distributions, alternative algorithms, such as Expectation-Maximization (EM), may be required. In other words, when only observed residues in the training data... and ψ j When one of the labels is available, the residue can be retained in the training data and the distribution can be adjusted based on the availability of the label. Decomposed into: when ψ j The tag is available when it is p( , and when The label is available when it is p(ψ) j | Because the marginal term p(φ) j Even when only φ is observed j Evaluation is also allowed during labeling, and for ψ j Similarly, the model can still learn from residues lacking a single angle label. In the example, residues with labels for two dihedral angles observed in the training data are assigned (e.g., randomly) to evaluate condition p(φ). j | ψ j ) or p(ψ j | φ j Residues of any one of the following, and only those with a single angle tag, are assigned to evaluate the edge term (p(φ)). j ) or p(ψ j )).

[0133] Generate pMHC conformation set Figure 7 Non-limiting examples are provided to demonstrate the process by which predicted distributions of pairwise distance data and dihedral angle data of pMHC complexes can be converted into structural constraints, and how a set of pMHC conformations can then be generated based on such structural constraints.

[0134] MHC sequences or pseudo-sequences (i.e., the ordered set of amino acids in an MHC molecule that typically contact the peptide), containing amino acid residues that typically contact the binding peptide, such as 34 amino acid residues (shown here as YFAMYGEKV…(SEQ ID NO:1)), and peptide sequences, such as, for example, 9-mer peptide sequences (shown here as LLFGYPVYV (SEQ ID NO:2)), are input into a trained machine learning model (e.g., Figure 1 The structural constraint prediction model 141 shown in the figure predicts pairwise distance data (shown here as the mean (702, left figure) and standard deviation (704, middle figure) values ​​characterizing the pairwise distance data distribution of all peptides with MHC pseudo-sequence amino acid residue pairs) and predicts dihedral angle data (shown here as the mean and standard deviation (706, right figure) values ​​of the peptide dihedral angle of all peptide residues in the complex).

[0135] like Figure 8 As shown in the figure, these predictions can then be used as structural modeling software (such as FlexPepDock (Rosetta, www.rosettacommons.orgThe constraints in the software package, along with the template structure model selected based on predicted pairwise distance and dihedral angle data, are used to generate a set of conformational isomers for the pMHC complex. The selected template structure (model 802, top left) is based on predicted pairwise distance and dihedral angle constraints applied to the pMHC crystal structure PDBID: 4NO5 and is used to create the initial model for conformational isomer generation using FlexPepDock (model 804, top right). Conformational isomer generation is then performed using the iterative process described elsewhere in this document. A superposition of 200 conformational isomer models is shown in model 806 below.

[0136] Generate a set of conformational isomers using a structural constraint prediction model. Figure 9 Non-limiting examples of a process 900 (e.g., a computer-implemented method) for generating multiple compatible structures of peptide-protein complexes are provided. Process 900 can be used... Figure 1 The prediction system 100 described herein is used for implementation. For example, process 200 can be implemented using... Figure 1 This is achieved using a sequence analyzer 108 and one or more machine learning models 132.

[0137] exist Figure 9 In step 902, peptide sequence data (of at least one peptide) and protein sequence data (of at least one protein) can be input into a trained machine learning model (e.g., Figure 1 The structural constraint prediction model 141 shown in the figure determines: (i) predicted pairwise distance data for at least one peptide amino acid residue of at least one peptide in a peptide-protein complex and at least one protein amino acid residue of at least one protein (e.g., each predicted pairwise distance data includes one or more predicted distances, and / or one or more parameters characterizing the predicted distribution of the distance between a pair of amino acid residues, which includes peptide residues and protein amino acid residues in the peptide-protein complex); and (ii) predicted dihedral angle data for at least one peptide amino acid residue of at least one peptide in a peptide-protein complex (e.g., each predicted dihedral angle data for a peptide residue in the peptide-protein complex includes one or more predicted angles for the ψ angle, One or more predicted angles of the angle, one or more parameters characterizing the predicted distribution of the angle ψ, characterizing One or more parameters of the predicted distribution of the angle, and / or characterizing ψ and One or more parameters of the joint distribution of the predicted angle.

[0138] In some instances, predicted pairwise distance data and predicted dihedral angle data may include a predicted posterior distribution for each parameter (e.g., each distance and each dihedral angle – or dihedral angle pair) to which predictions are made, which may be characterized, for example, by the mean, standard deviation, and uncertainty (or any other set of statistical parameters used to characterize the distribution). In other words, predicted pairwise distance data may include one or more prediction parameters that collectively characterize the distribution of pairwise distances between a pair of amino acid residues in a peptide-protein complex, the pair of amino acid residues comprising at least one peptide residue and at least one protein amino acid residue. Similarly, predicted dihedral angle data may include one or more parameters that collectively characterize the ψ-angle distribution of at least one amino acid residue in a peptide that forms part of a peptide-protein complex. Angular distribution and / or ψ and The joint distribution of angles. In some instances, the trained machine learning model can be further configured to determine the mean, standard deviation, and / or uncertainty (or any other set of statistical parameters used to characterize the distribution) of predicted pairwise distance data and / or predicted dihedral angle data. In an embodiment, predicted pairwise distance data is predicted for all pairs comprising peptide amino acid residues and protein pseudosequence residues, where protein pseudosequence residues are protein residues considered to be involved in the interaction with the peptide when forming a peptide-protein complex. In an embodiment, predicted dihedral angle data is predicted for all amino acid residues in the peptide except for the last (C-terminal) amino acid residue. In an embodiment, predicted dihedral angle data is predicted for all amino acid residues in the peptide core (also referred to as the minimal epitope in the context of a peptide-MHC complex).

[0139] In some instances, the input may include peptide sequence data of at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 20, 30, 40, 50, or more than 50 peptides or portions thereof. In some instances, the peptide sequence data may include sequence data of peptides with a length of at least 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 amino acid residues. In some instances, the peptide may include normal peptides (e.g., peptides expressed in normal cells) or portions thereof isolated from a subject (e.g., a patient), derived from nucleic acid sequences isolated from the subject, or obtained from one or more genomic, transcriptomic, and / or proteomic reference datasets (e.g., genomic, transcriptomic, or proteomic databases). In some instances, a peptide may include a mutant peptide (e.g., a neoantigen expressed or potentially expressed in tumor cells) or a portion thereof, isolated from or derived from a nucleic acid sequence isolated from a subject (e.g., a patient). In some instances, a peptide may include a bacterial peptide or a portion thereof expressed by bacteria. In some instances, a peptide may include a viral peptide or a portion thereof expressed in a virus. In some instances, a peptide may include a peptide derived from one or more pathogens, such as, for example, bacteria, viruses, parasites (e.g., Plasmodium), etc.

[0140] In some instances, peptide sequence data for at least one peptide may include peptide embedding representations (or peptide embeddings, e.g., machine-friendly vector representations of peptide sequences) of at least one peptide generated using a trained peptide embedding machine learning model. The vector representation of the peptide sequence can encode the peptide sequence based on statistical dependencies and patterns present in the peptide sequence used to train the model. The peptide sequence data may include peptide embedding representations from a trained protein language model, such as, for example, peptide embedding model 204. Peptide embedding model 204 may be trained in a self-supervised manner as part of a protein language model. Peptide embedding model 204 may be further trained (also referred to as “fine-tuning”) as part of a machine learning model configured to predict the properties of peptide-protein complexes containing peptides. Furthermore, multiple successive fine-tuning steps may have been implemented, including further training peptide embedding model 204 as part of machine learning models configured to predict different properties of peptide-protein complexes containing peptides. The properties of peptide-protein complexes may be selected from functional properties and structural properties. The machine learning model configured to predict the properties of peptide-protein complexes may include a peptide embedding model, a protein embedding model, and a task-specific prediction head. The task-specific prediction head may be a classification head or a regression head. The predicted functional property can be the binding between a peptide and a protein to form a peptide-protein complex. The binding between the peptide and protein can be predicted using a classification head as a binary property (binding / non-binding) or using a regression head as a continuous property (e.g., predicting binding affinity). The predicted functional property can be the presentation of the peptide by a protein on the cell surface, where the protein is an MHC molecule. The presentation of the peptide by an MHC molecule can be predicted as a binary property (e.g., presentation / non-presentation), for example, using eluted ligand data. The predicted functional property can be the stability of the peptide-protein complex. The stability of the peptide-protein complex can be predicted as a binary (e.g., stable / unstable) or continuous (e.g., measured complex half-life) property. The predicted structural property can be a prediction of pairwise distance data and / or dihedral angle data as described herein. The task-specific prediction head can be a multilayer perceptron. As understood by those skilled in the art, predicting binary properties can include predicting the probability that a peptide-protein complex containing the peptide belongs to a category in a pair of categories (e.g., predicting the probability of peptide binding to a protein). Peptide-MHC binding affinity data, binding data, and presentation data are available in public databases and datasets, such as those mentioned above by Chu et al. (2022), and the Immune Epitope Database and Tools (IEDB). www.iedb.org / home_v3.php ).

[0141] In some instances, the input may include protein sequence data of at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 100, 1000, or more than 1000 proteins or portions thereof. In some instances, the protein sequence data may include sequences or pseudo-sequences of at least 20, 22, 24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52, 54, 56, 58, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, or at least 200 amino acid residues. In some instances, the protein sequence data for at least one protein may include a protein embedding representation (or protein embedding, e.g., a machine-friendly vector representation of a protein sequence) generated using a trained protein embedding machine learning model. Vector representations of protein sequences can encode protein sequences based on statistical dependencies and patterns present in the protein sequences used to train the model. Protein sequence data can include protein embedding representations from a trained protein language model, such as, for example, protein embedding model 202. Protein embedding model 202 can be trained in a self-supervised manner as part of a protein language model. Protein embedding model 202 can be further trained (also known as “fine-tuning”) as part of a machine learning model configured to predict the properties of peptide-protein complexes containing proteins, as explained above regarding peptide embedding model 204.

[0142] In some instances, the trained peptide embedding machine learning model and the trained protein embedding machine learning model can be different models (e.g., respectively). Figure 2 The peptide embedding model 204 and protein embedding model 202 are shown in the example. In some instances, the trained peptide embedding machine learning model and the trained protein embedding machine learning model can be the same model (e.g., Figure 1 and Figure 2The embedding generation engine 140 shown in the example is illustrated. Exemplary methods for performing peptide and / or protein sequence embedding are described, for example, in Ofer et al. (2021), “The Language of Proteins: NLP, Machine Learning & Protein Sequences”, Computational and Structural Biotechnology Journal 19:1750-1758. In an embodiment, the trained peptide embedding machine learning model and the trained protein embedding machine learning model may be based on the same model obtained from the protein language model (i.e., the protein / peptide embedding model that has been trained in a self-supervised manner as part of the protein language model). Two instances of the protein / peptide embedding model (which may be referred to as the pre-trained protein embedding model and the pre-trained peptide embedding model, respectively) may have been included in a machine learning model trained to predict one or more properties of peptide-protein complexes to obtain the trained protein embedding model and the trained peptide embedding model (i.e., further training / fine-tuning the pre-trained protein embedding model and the peptide language model).

[0143] In some instances, trained peptide embedding machine learning models and / or trained protein embedding machine learning models can include large language models (LLMs), such as deep learning algorithms typically based on transformer networks, which can use very large datasets (e.g., very large datasets of peptide and / or protein sequences) to identify, summarize, translate, predict, and generate content. Large language models are described in more detail in, for example, Li et al. (2021), “BioSeq-BLM: APlatform for Analyzing DNA, RNA and Protein Sequences Based on BiologicalLanguage Models”, Nucleic Acids Res. 49, e129–e129, and Naveed et al. (2023), “A Comprehensive Overview of Large Language Models”, arxiv.org / pdf / 2307.06435.pdf. Protein language models can refer to large language models that have been trained in a self-supervised manner using data from one or more proteome databases. Therefore, a protein language model can be a large language model that has been trained to model the “language” of proteins using large amounts of unlabeled protein sequence data.

[0144] In some instances, trained peptide embedding machine learning models and / or trained protein embedding machine learning models may include evolutionary scale models (ESMs), such as transformer-based LLMs trained on unlabeled data of over 250 million distinct protein sequences to generate vector representations for input sequences. ESMs have been shown to encode the biochemical properties of amino acids in a sequence, biological variation and long-range homology of the sequence, as well as secondary structure and tertiary contacts of the sequence (see, for example, Rives et al. (2021), “Biological Structure and Function Emerge from Scaling Unsupervised Learning to 250 Million Protein Sequences”, PNAS 118(15)e2016239118). Trained peptide embedding machine learning models and / or trained protein embedding machine learning models may include fine-tuned versions of ESM models, where fine-tuning refers to further training the model as part of a machine learning model trained to predict one or more properties of peptide-protein complexes.

[0145] In some instances, a trained machine learning model is configured to determine the prediction of pairwise distance data and dihedral angle data (e.g., Figure 1 The structural constraint prediction model 141 shown herein may include a trained neural network, as described in more detail elsewhere herein. In some instances, the trained machine learning model may include, for example, a trained multilayer perceptron (MLP), a trained attention-based neural network, or a trained graph neural network.

[0146] In some instances, machine learning models are configured to determine predictions for pairwise distance data and dihedral data (e.g., Figure 1The structural constraint prediction model 141 shown can be trained by: receiving peptide sequence data of multiple peptides; receiving protein sequence data of multiple proteins; receiving peptide-protein binding data of multiple peptide-protein combinations; training a classifier (or regressor) based on the peptide sequence data, protein sequence data, and peptide-protein binding data to predict peptide-protein binding (or other functional properties of peptide-protein complexes) of specific combinations of peptides and proteins; and attaching a regression head to at least a portion of the trained classifier (or regressor) and training the regressor head based on pairwise distance data and dihedral angle data derived from crystal structure data of an exemplary set of peptide-protein complexes to predict the probability distribution of the pairwise distance data and dihedral angle data of peptide-protein complexes. The classifier can be a machine learning model including a protein embedding model, a peptide embedding model, and a classification head. In this context, the classification head can be a machine learning model configured to take peptide embedding representations from the peptide embedding model and protein embedding representations from the protein embedding model as input and provide a prediction of the category among multiple categories to which the peptide-protein complex belongs as output. Multiple categories may include categories of peptide-protein pairs that bind to each other to form a peptide-protein complex and categories of peptide-protein pairs that do not bind to each other to form a peptide-protein complex. The regressor may be a machine learning model comprising a protein embedding model, a peptide embedding model, and a regression head. In this context, the regression head may be a machine learning model configured to receive peptide embedding representations from the peptide embedding model and protein embedding representations from the protein embedding model as input, and to provide a prediction of the binding affinity between the protein and peptide as output (e.g., predicted binding affinity). Attaching the regressor head to at least a portion of a trained classifier / regressor and training the regressor head to predict the probability distribution of pairwise distance and dihedral angle data for peptide-protein complexes may include: removing the previous classifier / regressor head and replacing it with a new regressor head, and then training the resulting machine learning model to predict the probability distribution of pairwise distance and dihedral angle data for peptide-protein complexes based on pairwise distance and dihedral angle data derived from crystal structure data of an exemplary set of peptide-protein complexes. Crystal structure data may refer to pairwise distances and dihedral angles obtained from experimentally determined peptide-protein complex structural coordinates. Pair distances and dihedral angles can be calculated from structural coordinates obtained from publicly available structural databases, such as PDB.

[0147] In some instances, peptide sequence data for multiple peptides and / or protein sequence data for multiple proteins used for training can each independently include at least 2 x 10^6 pkbps. 4 4 x 10 4 6 x 10 4 8 x 10 4 1 x 105 2 x 10 5 3 x 10 5 4 x 10 5 5 x 10 5 6 x 10 5 7 x 10 5 8 x 10 5 9 x 10 5 1 x 10 6 2 x 10 6 3 x 10 6 4 x 10 6 5 x 10 6 6 x 10 6 7 x 10 6 8 x 10 6 9 x 10 6 1 x 10 7 or more than 1 x 10 7 Sequence data of a peptide or protein sequence.

[0148] In some instances, peptide-protein binding data for training multiple peptide-protein combinations may include at least 2 x 10-1 4 4 x 10 4 6 x 10 4 8 x 10 4 1 x 10 5 2 x 10 5 3 x 10 5 4 x 10 5 5 x 10 5 6 x 10 5 7 x 10 5 8 x 10 5 9 x 10 5 1 x 10 6 2 x 10 6 3 x 10 6 4 x 10 6 5 x 10 6 6 x 10 6 7 x 10 6 8 x 10 6 9 x 10 6 1 x 10 7 or more than 1 x 10 7 Binding data for individual peptide-protein combinations (e.g., binary (yes / no) binding data and / or continuous value binding affinity data).

[0149] In some instances, during the training of the regressor head, pairwise distance data and dihedral angle data derived from crystal structure data for each combination of a single peptide amino acid residue and a single protein amino acid residue in an exemplary peptide-protein complex can be presented as separate training instances. This provides more training data when the number of available peptide-protein crystal structures is limited. In other words, the training data used to train a machine learning model configured to determine predictive pairwise distance data and dihedral angle data (e.g., structure-constrained prediction model 141) for each of multiple training amino acid residue pairs containing peptide and protein amino acid residues in a training peptide-protein complex may include experimentally determined values ​​of the pairwise distance between the amino acid residues in that pair and one or two dihedral angles associated with the peptide residues in that pair. These pairwise distances and dihedral angles may be referred to herein as “grounded truth” distances / dihedral angles (or simply “grounded truth” or “tags”). Thus, the training data may include experimentally determined grounded truth pairwise distances and dihedral angles for each of multiple amino acid residue pairs derived from multiple training experiments determining the structure of the peptide-protein complex.

[0150] In some instances, the regressor can be further trained based on enhanced pairwise distance data and dihedral data, which can be derived by: receiving peptide sequence data and protein sequence data of at least one peptide and at least one protein known to bind and form a peptide-protein complex; processing the peptide sequence data and protein sequence data of at least one peptide and at least one protein using a previous version of a trained machine learning model to determine the uncertainty of the pairwise distance data and the uncertainty of the dihedral data of the peptide-protein complex; applying the determined uncertainty of the pairwise distance data and the determined uncertainty of the dihedral data to the structure of an exemplary peptide-protein complex to generate multiple compatible structures of the peptide-protein complex; and extracting the pairwise distance data and dihedral data of the multiple compatible peptide-protein complex structures to use as enhanced training data for the regressor. Processing peptide sequence data and protein sequence data of at least one peptide and at least one protein using a prior version of a trained machine learning model to determine uncertainties in pairwise distance data and dihedral data of peptide-protein complexes may include: obtaining pairwise distance data and dihedral data of peptide-protein complexes using a trained machine learning model (e.g., a trained machine learning model trained on training data without augmentation data), wherein the pairwise distance data includes statistical measures of the variability of pairwise distances (e.g., as one of a set of parameters characterizing the predicted distribution of pairwise distances between a pair of amino acid residues), and the dihedral data includes statistical measures of the variability of dihedral angles (e.g., as one of a set of parameters characterizing the predicted distribution of one or both of the dihedral angles of a pair of peptide amino acid residues), as described herein. Applying the determined uncertainties of the pairwise distance data and the determined uncertainties of the dihedral angle data to the structure of an exemplary peptide-protein complex to generate multiple compatible structures of the peptide-protein complex may include: generating a set of conformational isomers using the predicted uncertainty measures, structural modeling software, and a template structure as an experimentally determined structure of a particular peptide-protein complex. The resulting conformational isomers can then be used to derive enhanced "grounded truth" pairwise distances and dihedral angle sets. In other words, enhanced pairwise distance and dihedral angle data can be obtained by applying the methods described herein to peptide-protein complexes with experimentally determined structures to generate a set of conformational isomers, and using the resulting conformational isomers to derive additional true pairwise distances and dihedral angles of the peptide-protein complexes. These additional grounded truth pairwise distances and dihedral angles can capture a reasonably expected level of flexibility in the peptide-protein complexes that is not captured in experimentally determined crystal structures.In an embodiment, the regressor can be further trained using enhanced pairwise distance data and dihedral data, which can be derived by selecting one or more peptide-protein pairs known to bind to each other to form a complex and using a structure prediction machine learning algorithm to predict the structure of the complex corresponding to each of the peptide-protein pairs. For example, the machine learning algorithm can be selected from the AlphaFold model family (e.g., AlphaFold2 or a fine-tuned version thereof, such as AF-FT fine-tuned for peptide-HLA binding prediction, see Motmaen et al. Proc. Nat. Ac. Sci. 120(9), 2216697120, 2023).

[0151] In some instances, trained peptide embedding machine learning models and / or trained protein embedding machine learning models can be integrated with trained machine learning models and trained simultaneously with them.

[0152] exist Figure 9 In step 904, the initial structure of the peptide-protein complex (also referred to as the “template” in this paper) can be identified based on predicted pairwise distance data and predicted dihedral angle data.

[0153] Identifying the initial structure of a peptide-protein complex based on predicted pairwise distance data and predicted dihedral angle data (referred to herein as the “initial template structure” or simply the “template structure”) may include, for example: obtaining multiple crystal structures of the peptide-protein complex from a protein structure database; and selecting a crystal structure from the multiple crystal structures based on the similarity between the pairwise distance data and dihedral angle data determined for the crystal structure and the predicted distance data and predicted dihedral angle data. The similarity between the pairwise distance data and dihedral angle data determined for the crystal structure and the predicted distance data and predicted dihedral angle data can be quantified using a likelihood estimate of the pairwise distance data and dihedral angle data determined for the crystal structure, under a distribution characterized by the predicted distance data and predicted dihedral angle data. Alternatively, a structural similarity metric determined for the crystal structure can be used, along with a metric based on a trained model (e.g., Figure 1The structure constraint prediction model 141 shown in the diagram determines the maximum likelihood estimate of the predicted pairwise distance and dihedral angle based on the predicted pairwise distance and dihedral angle data, and selects a crystal structure from multiple crystal structures. In an embodiment, the initial structure is selected based on the sequence similarity between the peptide in the peptide-protein complex and peptides of the same length in the crystal structure of the peptide-protein complex. Sequence similarity can be calculated using any scoring known in the art, such as, for example, BLOSUM62. The multiple crystal structures obtained can be crystal structures containing the same protein as the peptide-protein complex. For example, one or more crystal structures can be used for a specific MHC molecule. The multiple crystal structures obtained can be crystal structures containing the same protein as the peptide-protein complex, but not necessarily the same peptide. The multiple crystal structures obtained can be crystal structures containing peptides of the same length as the peptide in the peptide-protein complex to be modeled. The multiple crystal structures can be obtained by clustering crystal structures containing peptides of the same length as the peptide in the peptide-protein complex to be modeled and identifying representative structures in each cluster. In an embodiment, the initial structure (initial template structure) can include multiple structures. For peptides up to 10 amino acids, a single template structure can be used. For peptides of 11 or more amino acids, multiple template structures can be used. Multiple template structures can include crystal structures obtained by clustering crystal structures containing peptides of the same length as the peptides in the peptide-protein complex to be modeled and identifying representative structures for each cluster.

[0154] In some instances, multiple crystal structures of peptide-protein complexes obtained from protein structure databases may include at least 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 120, 140, 160, 180, 200, 220, 240, 260, 280, 300, or more than 300 crystal structures.

[0155] In some instances, protein structure databases may include, for example, protein databases (PDBs) (ResearchCollaboratory for Structural Bioinformatics; rcsb.org).

[0156] exist Figure 9 In step 906, multiple compatible structures of the peptide-protein complex can be generated based on the initial structure, predicted pairwise distance data, and predicted dihedral angle data.

[0157] In some instances, each compatible structure of the peptide-protein complex may, for example, correspond to the posterior distribution of predicted pairwise distance data and the posterior distribution of predicted dihedral angle data output by a trained machine learning model. In some instances, each compatible structure of the peptide-protein complex may correspond to sequence data of at least one peptide and at least one protein, as well as to the predicted pairwise distance data and the predicted dihedral angle data. In embodiments, multiple compatible structures of the peptide-protein complex are used to obtain predictions of at least one auxiliary feature.

[0158] In some instances, generating multiple compatible structures of a peptide-protein complex based on an initial structure, predicted pairwise distance data, and predicted dihedral angle data may include, for example, inputting the initial structure of the peptide-protein complex into a structure modeling software package (e.g., Figure 1 and Figure 2 The ensemble generation engine 142 shown in the document inputs peptide sequence data, predicted pairwise distance data, and predicted dihedral angle data determined by a trained machine learning model for peptide-protein complexes into a structure modeling software package; and iteratively samples and evaluates a structure scoring function from multiple possible structures of the peptide-protein complex to identify multiple compatible structures of the peptide-protein complex, wherein a compatible structure of the peptide-protein complex is a structure that: (i) is compatible with one or more constraints imposed by the predicted pairwise distance data and the predicted dihedral angle data, and (ii) optimizes the structure scoring function. The input peptide sequence data is optional, and conversely, the initial structure may have been adapted to include peptide sequences, such as initial structures using, for example, homology modeling and containing peptides with sequences different from those in the peptide-protein complex to be modeled.

[0159] In some instances, iterative sampling may include 2, 4, 6, 8, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, or more than 1000 iterations. The number of iterations can be set by one or more stopping criteria, such as the MCMC algorithm. Therefore, the number of iterations may not be preset and may vary between different instances.

[0160] In some instances, the plurality of possible structures may include 2, 4, 6, 8, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, or more than 1000 possible structures. In embodiments, the plurality of possible structures includes structures between 10 and 1000, structures between 50 and 1000, structures between 50 and 500, structures between 100 and 500, structures between 100 and 400, or approximately 200 structures.

[0161] In some instances, iterative sampling from multiple possible structures of a peptide-protein complex can be performed using Markov Chain Monte Carlo (MCMC) sampling algorithms (see, for example, Ravenzwaaij et al. (2018) “A Simple Introduction to Markov Chain Monte–Carlo Sampling”, Psychon Bull Rev25:143–154, and Jones et al. (2022), “Markov Chain Monte Carlo in Practice”, Annu. Rev.Stat.Appl. 9:557–578). In some instances, Markov Chain Monte Carlo (MCMC) sampling algorithms may include, for example, the Metropolis-Hastins algorithm, the Hamilton Monte Carlo algorithm, or the Langevin Monte Carlo algorithm. In the embodiments, iterative sampling from multiple possible structures of the peptide-protein complex can be performed using the MCMC algorithm with the Metropolis standard, as described, for example, in Raveh et al., Proteins, Vol. 78, No. 9, July 2010, pp. 2029-2040.

[0162] In some instances, the structure scoring function may include, for example, the free energy calculation of the peptide-protein complex (e.g., the binding free energy, which contains the free energy difference between the bound complex and the unbound peptide and protein; or the total interaction free energy, which contains the free energy difference between the bound complex and the separated peptide and protein). For example, the structure scoring function may be as described in Raveh et al., Proteins, Vol. 78, No. 9, July 2010, pp. 2029-2040, or the Rosetta scoring function (see Leman et al., Nature Methods 17(7), 665-680(2020)). In some instances, the structure scoring function may include a proxy for the free energy calculation. In some instances, the structure scoring function may include a proxy for the free energy calculation of peptide-protein binding, a proxy for the free energy calculation of at least one peptide-protein interface feature, or any combination thereof.

[0163] In some instances, the structure scoring function (e.g., a proxy for free energy calculations) can be based on one or more of the following: ab initio quantum mechanical calculations, density functional theory (DFT) calculations, semi-empirical calculations, molecular mechanical force field calculations, statistical potential calculations, neural potential (or neural force field) calculations predicting forces acting on individual atoms (e.g., using models that reproduce the results of advanced quantum mechanical calculations at a higher speed and / or provide more robust assessments of poor structures), or a function parameterized by a machine learning model trained on structure data. For example, in some cases, the structure scoring function can be a physics-based energy function that determines the total energy of a peptide-protein complex, including one or more of electrostatic energies, covalent bond energies, van der Waals energies, etc. It should be understood that different structure scoring functions can have different levels of accuracy and computational complexity. For example, a more accurate structure scoring function, such as one based on ab initio quantum mechanics, may incur greater computational overhead than a less accurate structure scoring function, such as one based on molecular mechanical force fields. Therefore, in some cases, more than one structure scoring function can be applied as part of the identification of compatible structures of peptide-protein complexes, wherein compatible structures are identified as those with, for example, the lowest structure score (e.g., the lowest total energy).

[0164] In some instances, structural modeling software packages may include, for example, AlphaFold (DeepMind, London, UK), Amber (ambermd.org), ESMFold (Meta AI, New York City, NY), FlexPepDock (RosettaCommons, rosettacommons.org), HADDOCK (High Ambiguity Drivenbiomolecular DOCKing; bonvinlab.org), Molecular Operating Environment (MOE; Chemical Computing Group, Montreal, Canada), OpenMM (openmm.org), SchrödingerGlide (New York City, NY), and AutoDock (autodock.scripps.edu). In this embodiment, the structural modeling software package includes FlexPepDock. Structural modeling software may also be referred to as a "physics-based engine" because it determines an energy-reasonable structure based on a predetermined energy function. Any structural modeling algorithm that can obtain a predicted structure using a template sequence and can accommodate dihedral angles and / or pairwise distance constraints (e.g., within an energy scoring function) can be used in the context of this disclosure.

[0165] In some instances, Figure 9 The process (or computer-implemented method) shown may further include predicting at least one auxiliary feature of the peptide-protein complex based on multiple compatible structures of the peptide-protein complex. In some instances, the at least one auxiliary feature may include, for example, prediction of the binding of at least one peptide to at least one other protein or protein complex, and / or prediction of at least one interface feature of the peptide-protein complex (such as, for example, information identifying complementary interface features between the peptide-protein complex and another protein or protein complex, and / or information identifying complementary surface fingerprint features between the peptide-protein complex and another protein or protein complex).

[0166] In some instances, Figure 9 The process (or computer-implemented method) shown may further include displaying a subset of predicted pairwise distance data and / or predicted dihedral angle data using any of the various techniques known to those skilled in the art for displaying three-dimensional tensor data (e.g., 3D graphs, 2D graphs of tensor data specifying slices, etc.). In some cases, the method may include displaying the subset of predicted pairwise distance data determined by a trained machine learning model as, for example, a two-dimensional heatmap containing the α-carbon (C) of peptide amino acid residues. α) and the α-carbon (C) of protein amino acid residues α A plot of pairwise distances between peptide and protein amino acid residues as a function of their respective positions. The plotted pairwise distances may include maximum likelihood estimates of the predicted pairwise distances, such as prediction parameters based on the corresponding distributions of the pairwise distances. The plotted pairwise distances may include the predicted mean of the predicted distributions of the pairwise distances. In other words, the predicted pairwise distance data may include parameters characterizing the distribution of each pairwise distance for its predicted pairwise distance data. These parameters can be used to obtain maximum likelihood estimates of each of these pairwise distances. The maximum likelihood estimate may be, for example, the predicted mean (i.e., the mean). This may be the case, for example, when predicting parameters of a Gaussian distribution (or any other distribution where the mean is the highest likelihood value of the distribution). The plotted pairwise distances may include multiple corresponding two-dimensional plots, each showing the predicted values ​​of the parameters of the predicted distributions of the pairwise distances. For example, the plotted pairwise distances may include a first two-dimensional plot showing the mean of the predicted distributions of the pairwise distances, and a second two-dimensional plot showing the standard deviation or other variability statistics of the predicted distributions of the pairwise distances. In some instances, the method may include displaying at least a subset of predicted dihedral data determined by a trained machine learning model as, for example, at least a subset of peptide amino acid residues. Two-dimensional scatter plot of the dihedral angle ψ. A two-dimensional scatter plot of the ψ dihedral angle can include each of the multiple peptide amino acid residues. Indicators of the mean or maximum likelihood estimate of the predicted distribution of the dihedral angle ψ. A two-dimensional scatter plot of the ψ dihedral angle can include each of the multiple peptide amino acid residues. The statistical measure of variability (e.g., standard deviation) of the predicted distribution of the ψ dihedral angle. Alternatives or exceptions In addition to the two-dimensional scatter plot of the ψ dihedral angle, it is possible to draw One or both of the scatter plot of the angle and / or the scatter plot of the ψ dihedral angle. Each scatter plot may include each of the multiple peptide amino acid residues. Indicators of the mean or maximum likelihood estimate of the predicted distribution of the ψ dihedral angle, and / or each of the multiple peptide amino acid residues. Indicators of the statistical measures of variability in the predicted distribution of the dihedral angle ψ.

[0167] In some instances, Figure 9The process (or computer-implemented method) described herein may further include iteratively traversing multiple candidate peptides (e.g., at least 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 30, 40, 50, 60, 70, 80, 90, 100 or more than 100 candidate peptide sequences or portions thereof) to identify at least one peptide having the highest likelihood for forming a peptide-protein complex.

[0168] In some instances, Figure 9 The process (or computer-implemented method) illustrated may further include predicting the characteristics of the peptide-protein complex (e.g., interface features, surface complementarity features, etc.) based on multiple compatible structures of the peptide-protein complex. In some instances, predicting the binding of the peptide-protein complex based on multiple compatible structures may include, for example: identifying complementary interface features between the peptide-protein complex and another protein or protein complex; identifying complementary surface fingerprint features between the peptide-protein complex and another protein or protein complex; identifying latent representations of the peptide-protein complex and another protein or protein complex; and providing the complementary interface features, complementary surface fingerprint features, and latent representations of the peptide-protein complex and another protein or protein complex as input to a second trained machine learning model configured to output a prediction of the binding of the peptide-protein complex to another protein or protein complex. The latent representation of the peptide-protein complex can be any representation in the space to which the peptide-protein complex is projected. This can be used, for example, to identify the similarity between complexes in the latent space (e.g., identifying complexes that are more similar to each other). In some instances, predicting the binding of peptide-protein complexes based on multiple compatible structures of the peptide-protein complex can include, for example, predicting the structure of multiple compatible structures complexed with another protein.

[0169] In some instances, complementary interface features may include, for example, a paired list of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, a paired list of interaction energies of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, the total hydrogen bond energy of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, a paired list of hydrogen bond energies of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, the total van der Waals interaction energy of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, a paired list of van der Waals interaction energies of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, or any combination thereof.

[0170] In some instances, complementary surface fingerprint features may include, for example, complementary shape features based on steric hindrance, complementary electrostatic features, total electrostatic interaction energies of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, a paired list of electrostatic interaction energies of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, total cation-π interaction energies of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, a paired list of cation-π interaction energies of amino acid residues in a peptide-protein complex that are in contact with amino acid residues in another protein or a protein complex, complementary polarity features, or any combination thereof.

[0171] In some instances, complementary surface fingerprint features between a peptide-protein complex and another protein or protein complex can be predicted by a third trained machine learning model based on the potential representations of the peptide-protein complex and another protein or protein complex.

[0172] In some instances, the second trained machine learning model and / or the third trained machine learning model may include, for example, a trained neural network or a deep learning model.

[0173] Train a classifier to predict pairwise distance data and dihedral data. Figure 10 Provides methods for training machine learning models (e.g., Figure 1 The flowchart shown in the image is a non-limiting example of a process 1000 (e.g., a computer-implemented method) for predicting the distribution of pairwise distance data and dihedral angle data of peptide-protein complexes using a structure-constrained prediction model 141. Process 1000 can be used... Figure 1The prediction system 100 described herein is used to implement this. For example, process 700 can be implemented using... Figure 1 This is achieved using a sequence analyzer 108 and one or more machine learning models 132. All the features described above, for example with respect to Figure 9, can be equally applied to the method of Figure 10.

[0174] exist Figure 10 In step 1002, peptide sequence data of multiple peptides can be received (e.g., from...). Figure 1 (One or more processors of the prediction system 100 described herein).

[0175] In some instances, peptide sequence data for multiple peptides may include at least 2 x 10^6 peptides. 4 4 x 10 4 6 x 10 4 8x10 4 1 x 10 5 2 x 10 5 3 x 10 5 4 x 10 5 5 x 10 5 6 x 10 5 7 x 10 5 8 x 10 5 9 x 10 5 1x10 6 2 x 10 6 3 x 10 6 4 x 10 6 5 x 10 6 6 x 10 6 7 x 10 6 8 x 10 6 9 x 10 6 1 x 10 7 or more than 1 x 10 7 Sequence data of one peptide sequence.

[0176] In some instances, peptide sequence data for multiple peptides may include peptide embedding representation data for multiple peptides (e.g., machine-friendly vector representations of peptide sequences), which are generated using a trained peptide embedding machine learning model (e.g., ...). Figure 1 The embedded engine 140 shown in the document generates the data as described elsewhere in this document.

[0177] exist Figure 10 In step 1004, protein sequence data of multiple proteins can be received.

[0178] In some instances, protein sequence data for multiple peptides may include at least 2 x 10^6 peptides. 4 4 x 10 4 6 x 10 4 8 x 10 4 1 x 10 5 2 x 10 5 3 x 10 5 4 x 10 5 5 x 10 5 6 x 10 5 7 x 10 5 8 x 10 5 9 x 10 5 1 x 10 6 2 x 10 6 3 x 10 6 4 x 10 6 5 x 10 6 6 x 10 6 7 x 10 6 8 x 10 6 9 x 10 6 1 x 10 7 or more than 1 x 10 7 Sequence data of a protein sequence.

[0179] In some instances, protein sequence data for multiple proteins may include protein embedding representation data for multiple proteins (e.g., machine-friendly vector representations of protein sequences), which are generated using a trained protein embedding machine learning model (e.g., Figure 1 The embedding engine 140 shown in the example generates the model, as described elsewhere in this document. In some instances, the trained protein embedding machine learning model can be identical to the trained peptide embedding machine learning model.

[0180] exist Figure 10 In step 1006, peptide-protein binding affinity data for multiple peptide-protein combinations can be received.

[0181] In some instances, peptide-protein binding data for multiple peptide-protein combinations may include at least 2 x 10-1 4 4 x 10 4 6 x 10 4 8 x 10 4 1 x 10 5 2 x 10 5 3 x 10 5 4 x 105 5 x 10 5 6 x 10 5 7 x 10 5 8 x 10 5 9 x 10 5 1 x 10 6 2 x 10 6 3 x 10 6 4 x 10 6 5 x 10 6 6 x 10 6 7 x 10 6 8 x 10 6 9 x 10 6 1 x 10 7 or more than 1 x 10 7 Binding data for individual peptide-protein combinations (e.g., binary (yes / no) binding data and / or continuous value binding affinity data).

[0182] exist Figure 10 In step 1008, a classifier or regressor can be trained based on peptide sequence data, protein sequence data, and peptide-protein binding data (e.g., Figure 1 The structurally constrained prediction model 141 shown in the diagram is used to predict peptide-protein binding for specific combinations of peptides and proteins. In other words, a classifier or regressor can be trained using training data to predict peptide-protein binding data for input peptide sequence data and protein sequence data pairs. This training data, for each of several training peptide-protein complexes, includes: peptide sequence data, protein sequence data, and corresponding ground truth (i.e., measured or otherwise known) peptide-protein binding data. The ground truth peptide-protein binding data may include binary classification labels (e.g., bound / not bound) used to train the classifier, and / or a measure of binding affinity, which may be used to train the regressor or from which binary labels (e.g., for training the classifier) ​​may be obtained using a predetermined binding affinity threshold.

[0183] Any combination of machine learning methods and algorithms (wherein, as referred to herein, a machine learning model includes a trained machine learning algorithm) can be used to implement the disclosed methods. For example, a machine learning model can include a supervised learning model (i.e., a model trained using a labeled training dataset), an unsupervised learning model (i.e., a model trained using an unlabeled training dataset), a semi-supervised learning model (i.e., a model trained using a combination of labeled and unlabeled training data), a deep learning model (e.g., a neural network containing multiple layers of coupled "nodes," which can be trained in a supervised, unsupervised, or semi-supervised manner), or any combination thereof. For example, a machine learning model can include one or more machine learning models (such as, for example, embedding models) that have been pre-trained in a self-supervised manner and then further trained in a supervised or semi-supervised manner (e.g., as part of a corresponding machine learning model trained using corresponding training data including at least some labeled data to perform a corresponding prediction task). Embedding models, classifier heads, and / or regression heads can each be deep learning models. Therefore, a machine learning model including any or more of these can also be a deep learning model. A deep learning model can be an artificial neural network (ANN) comprising multiple hidden layers. Deep learning models may include one or more attention-based layers, one or more fully connected layers, one or more convolutional layers, and combinations thereof. In some instances, the disclosed methods may be implemented using one or more machine learning models (e.g., 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, or more than 10 machine learning models) or a set thereof. In some instances, the disclosed methods may be implemented using a continuous learning approach, wherein the machine learning models may be periodically or continuously updated based on new training data provided by, for example, a single local operating system, multiple local operating systems, or multiple geographically distributed operating systems.

[0184] Examples of machine learning algorithms that can be used include, but are not limited to: neural networks, feedforward neural networks (also known as multilayer perceptrons), recurrent neural networks, convolutional neural networks, deep neural networks, deep feedforward neural networks, deep recurrent neural networks, deep convolutional neural networks, attention-based neural networks, graph neural networks, or any combination thereof.

[0185] Neural networks (NNs) typically consist of interconnected groups of nodes organized into multiple layers. For example, an NN architecture may include at least an input layer, one or more hidden layers, and an output layer. NNs may include any total number of layers (e.g., 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 30, 40, 50, 60, 70, 80, 90, 100, or more than 100 layers) and any number of hidden layers (e.g., 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20, 30, 40, 50, 60, 70, 80, 90, 100, or more than 100 layers), where the hidden layers serve as trainable feature extractors, allowing the input dataset to be mapped to predicted output values ​​or sets of output values. Each layer of a neural network comprises multiple nodes (e.g., at least 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 25, 50, 75, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10,000, or more than 10,000 nodes). A node receives input data (e.g., peptide sequence data, protein sequence data, their concatenation, or other types of input data) directly from one or more input data nodes or from the outputs of one or more nodes in a previous layer, and performs a specific operation, such as a summation operation. In some cases, the connections from the input to the node are related to weights (or weighting factors). In some cases, for example, a node can, for example, take all inputs X... i and related weights W i The sum of the products. In some cases, the weighted sum is offset by a bias b. In some cases, the node output can be gated using a threshold or an activation function f, where f can be a linear or nonlinear function. The activation function can be, for example, a rectified linear unit (ReLU) activation function or other functions such as the saturated hyperbolic tangent function, identity function, binary step function, logistic function, arctangent function, softsign function, parametric rectified linear unit, exponential linear unit, softPlus function, bending identity function, softExponential function, sine wave function, sine function, Gaussian function or sigmoid function, or any combination thereof.

[0186] The weighting factors, biases, thresholds, or other computed parameters of a neural network can be "taught" or "learned" during the training phase using one or more training datasets (e.g., 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, or more than 10 training datasets). For example, the parameters can be trained using input data from the training datasets and gradient descent or backpropagation methods such that the output values ​​computed by the ANN (e.g., predicting pairwise distance data and / or dihedral data) are consistent with the examples included in the training datasets. The adjustable parameters of the model can be obtained using a backpropagation neural network training process, which may or may not use the same hardware used to implement the trained model and process the input peptide and / or protein sequence data.

[0187] Accessing the training dataset may include, for example, retrieving the training dataset from local or remote storage, loading the training dataset, and / or requesting (and receiving) part or all of the training dataset from one or more data stores (e.g., cloud data storage, server systems, or other data sources).

[0188] Training datasets can be randomly parsed, shuffled, and / or partitioned to train classifier models. Machine learning models can be trained using static or dynamic learning rates. Dynamic learning rates can be generated using, for example, learning rate annealing (e.g., using stepwise annealing or cosine annealing). Training can be performed using, for example, classification loss functions and / or regression loss functions. Loss functions can be based on, for example, mean squared error, median squared error, mean absolute error, median absolute error, entropy-based error, cross-entropy error, and / or binary cross-entropy error. Validation data (e.g., a separated subset of the training dataset used to train the machine learning model) can be used to evaluate the performance of the machine learning model during training. Training can be terminated if and / or when the target performance is achieved and / or the maximum number of training iterations has been completed.

[0189] exist Figure 10In step 1010, a regressor head may be attached to at least a portion of a trained classifier, and the model comprising the regressor head and that portion of the trained classifier may be trained based on pairwise distance data and dihedral data derived from crystal structure data of an exemplary peptide-protein complex set to predict the probability distributions of pairwise distance data and dihedral data for peptide-protein complexes. Therefore, a machine learning model comprising at least a portion of a regressor head and a trained classifier (e.g., comprising a trained protein embedding model and a trained peptide embedding model) may be trained using training data, which for each of a plurality of training peptide-protein complexes includes: protein sequence data, peptide sequence data, and corresponding ground truth structure data (e.g., experimentally determined crystal structure data) or pairwise distance data and dihedral data derived therefrom. This portion of the trained classifier may include the protein embedding model and the peptide embedding model as described herein, which are trained as part of a classifier model that further includes a classifier head, but which is excluded after such training.

[0190] As described elsewhere in this document, the predicted pairwise distance data and the predicted dihedral angle data may include a predicted posterior distribution for each parameter (e.g., for each pairwise distance and each dihedral angle to which predictions are made), which may be characterized, for example, by the mean, standard deviation, and uncertainty (or any other set of statistical parameters characterizing the distribution). In some instances, the trained machine learning model may be further configured to determine the mean, standard deviation, and / or uncertainty (or any other set of statistical parameters characterizing the distribution) for the predicted pairwise distance data and / or the predicted dihedral angle data.

[0191] In some instances, the exemplary set of peptide-protein complexes may include at least 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 120, 140, 160, 180, 200, 220, 240, 260, 280, 300, or more than 300 peptide-protein complexes.

[0192] As described elsewhere in this document, in some instances, during the training of the regressor head, pairwise distance data and dihedral angle data derived from crystal structure data for each combination of individual peptide amino acid residues and individual protein amino acid residues in an exemplary peptide-protein complex can be presented as separate training instances, thereby providing more training data when the number of available peptide-protein crystal structures is limited.

[0193] In some instances, the regressor can be further trained based on enhanced pairwise distance data and dihedral data, which can be derived by: receiving peptide sequence data and protein sequence data of at least one peptide and at least one protein known to bind and form a peptide-protein complex; processing the peptide sequence data and protein sequence data of at least one peptide and at least one protein using a previous version of a trained machine learning model to determine the uncertainty of the pairwise distance data and the uncertainty of the dihedral data of the peptide-protein complex; applying the determined uncertainty of the pairwise distance data and the determined uncertainty of the dihedral data to the structure of an exemplary peptide-protein complex to generate multiple compatible structures of the peptide-protein complex; and extracting the pairwise distance data and dihedral data of the multiple compatible peptide-protein complex structures to use as enhanced training data for the regressor.

[0194] In some instances, trained peptide embedding machine learning models and / or trained protein embedding machine learning models can be integrated with trained machine learning models and trained simultaneously with them.

[0195] As described above, the training dataset can be randomly parsed, shuffled, and / or partitioned to train the regressor head. The loss function used as part of the model training can include an error term (e.g., mean squared error or median squared error) and / or an entropy term (e.g., cross-entropy or binary cross-entropy).

[0196] Static or non-static learning rates can be used during training of the regressor head. For example, learning rate annealing (e.g., using stepwise annealing or cosine annealing) can be used to reduce the learning rate over multiple training iterations. In some instances, validation data evaluation can be used to potentially terminate training earlier (e.g., when it is determined that a performance target has been reached).

[0197] Example – Evaluation of Predictive Structural Models Human immune surveillance mechanisms are driven by CD8+ T cells, which recognize cytoplasmic peptides presented by human leukocyte antigen (HLA) molecules via their receptors. Understanding and predicting the specificity of T cell receptors (TCRs) for their homologous antigens is particularly challenging, and this process is further complicated by the heterogeneity of TCRs. TCR recognition of peptide-HLA (pHLA) molecules has also been attributed to the conformational flexibility and dynamics of peptides. This is especially true for complexes undergoing structural silencing mutations, where the co-crystal structure does not exhibit significant conformational changes. Despite these structural silencing mutations, TCRs recognizing such peptides still exhibit high specificity. For example, both wild-type (WT) and mutant (MT) PIK3CA peptides are derived from HLA-A. Presented at 03:01, but only the MT peptide is recognized by TCR4. To study such systems, traditional molecular dynamics (MD) simulations on long timescales can be employed. While MD works for a small number of instances, they can be limiting when large-scale data generation is required, especially when solving problems involving TCR-pHLA binding and / or signal transduction prediction. Recent work has utilized machine learning (ML) models to predict protein conformations. For example, AlphaFold2 (AF2) (Jumper, JM et al. PLOS Computational Biology 14(12), 1006342 (2018)) predicts static structures as seen in the PDB. AlphaFold-FineTune (AF-FT) (Motmaen, A, et al., PNAS, 120(9), 2216697120 (2023)) was specifically developed for the pHLA space. It fine-tunes AF2 for pHLA binding prediction but does not focus on modeling the conformational flexibility of the peptide. Despite considerable progress in ML-assisted conformational sampling, dedicated ML methods for generating peptide conformational isomers at pHLA interfaces remain underexplored. Evaluation of predicted conformational isomer distributions can also benefit from more detailed analysis, particularly when compared with MD (molecular dynamics) distributions. The inventors have developed a method using ML to guide a physics-based engine (such as Rosetta (Leman et al., Nature Methods 17(7), 665–680 (2020))) to model the conformational flexibility of peptides at pHLA-I interfaces. The ML component of this method is trained based on pHLA sequences and structures and predicts structural descriptors such as the distance between HLA and peptide residues and the dihedral angle of the peptide backbone. The predicted structural descriptors are applied as constraints in the physics engine to peptide docking schemes (Rosetta's FlexPepDock (FPD) scheme is used as an example here; Raveh et al. Proteins: Structure, Function, and Bioinformatics 78(9), 2029-2040, (2010)) to sample peptide conformations. The inventors tested this method on a pHLA complex benchmark set consisting of 32 different alleles spanning 8-13 amino acids. They demonstrated that this method (referred to as "p-flex") can recover 78% and 67% of the peptide backbone sampled from molecular dynamics simulations of peptides with lengths of 8-10 and 11-13 amino acids, respectively.To the best of their knowledge, this is the first method to capture the flexibility of peptides up to 13 amino acids in length within minutes without relying on time-consuming MD simulations.

[0198] Non-limiting examples of the methods and results disclosed herein are presented herein. An ML model is designed to predict the distribution of structural descriptors, namely the Cα-Cα distance between HLA and peptide residues and the dihedral angles (φ and ψ) of the peptide backbone. First, a classifier is trained on sequence data with pHLA binding tags to learn the association between peptides and HLA. Then, a regressor head from a common embedding shared with the classifier is trained based on structural descriptors of pHLA complexes from a protein database (PDB). The regressor outputs parameters controlling the distribution of distances between peptides and HLA residues and the joint distribution of the φ and ψ dihedral angles of the peptide residues. During inference, the model receives peptide and HLA sequences and outputs parameters controlling the predicted distribution of structural descriptors. The predicted descriptors are used to select template structures from the PDB similar to a given pHLA complex to initialize the FlexPepDock scheme (FPD). During refinement, they are also passed to the FPD as harmonic constraints and cyclic harmonic constraints. Although the model primarily learns these descriptors from static eutectic structures in the training set, the inventors argue that a well-regularized probabilistic model can extract shared information from different complexes with similar structural properties—not only about static structure but also about dynamics. The example results below support this claim; our model, trained on static structures, can recover the static structural descriptors of unseen test complexes, as well as approximate their dynamic distribution estimated by MD simulations.

[0199] Training Data – 3D Structures: The 3D structure data used for training and validating the ML method was compiled from PDBs. All PDB structures with the keyword “HLA” or “HLA”, restricted to “Homo sapiens” as the source organism, and with a resolution of 10 Å or lower were downloaded (as of February 2024). Each chain in the downloaded PDB structures was trimmed to a length of 180 amino acids. Chains with a length (i) less than 15 amino acids were labeled as peptides, and (ii) chains within the template neighborhood of 180 residues and genotyped by sequence alignment with HLA sequences in the IMGT / IPD database were labeled as HLA. If multiple assemblies were available in a given PDB, HLA chains and corresponding peptide chains were identified by measuring the distance between the second residue of the peptide and the 65th residue of the HLA, thus selecting a single pHLA pair (these residues were arbitrarily selected based on manual examination of multiple crystal structures in the PDB). The final trimmed snapshot of the pHLA structures was saved in the structure dataset. The genotyping alleles, along with the amino acid sequences of the pruned mid-chain, were stored in a separate metadata file. Structures of pHLA in both unbound and bound (with TCR or co-receptor) conformations were preserved, but the binding chaperone was removed before training. The final dataset contains a total of 651 noisy (called noisy because they include structures with deletion densities, bound, and unbound pHLA) structures, which were subdivided into training, validation, and test (blind) sets. PDB contains pHLA structures up to 10 amino acids in length (592 complexes) compared to those with lengths of 11–13 amino acids. To overcome the data sparsity of longer peptides, the inventors augmented the training set with structures predicted by AF-FT (Motmaen et al., PNAS 120(9), 2216697120 (2023)), which is an extension of AF2 (Jumper et al., PLOS Computational Biology 14(12), 1006342 (2018)) fine-tuned on the pHLA structures in the PDB. Specifically, to generate augmented structural data for peptides of 10–14 amino acids in length, the inventors randomly sampled 200 conjugates from the pHLA-I sequence obtained by Chu. et al. (Nature Machine Intelligence 4(3):300–311) and modeled them using default weights via Alphafold-finetune. Each model was then refined three times in a Rosetta force field using the Relax scheme.The regressor component of the method is then trained using structural descriptors (distance and dihedral) from a set of 3000 models.

[0200] Training Data - Labeled pHLA-I Sequences: A consolidated dataset of pHLA-I sequences with binary binding tags was obtained from Chu et al. (Nature Machine Intelligence 4(3):300-311). This dataset consists of (i) binding pHLA sequences and (ii) non-binding pairs, generated by randomly selecting peptides of the same length from the peptide groups of the same alleles in IEDB for each binding pair. In this labeled dataset, 2.2M sequence pairs are distributed across 112 unique HLAs, with each HLA representing approximately 17% of the binding peptide samples.

[0201] ML methods for predicting structural descriptors: The input to the constrained prediction model consists of HLA pseudo-sequences of a fixed length of 34 residues. HLA And peptide sequences of variable length between 8 and 15 residues (inclusive) 肽 Composition. In this context, the model is trained to output the predicted distribution of pairwise distances between peptides and HLA residues, as well as the φ and ψ dihedral angles of the peptide residues. As described in this article, it is represented as P(D, , ψ| s 肽 , s HLA ), where in this case the pairwise distance matrix is (Because a 34-amino acid HLA pseudo-sequence was used). This is called "joint" modeling of the dihedral angle, and this method is used unless otherwise stated. In practice, the distribution is broken down between residues, but for a given residue, φ and ψ are jointly modeled. The predicted distribution is parameterized as a Gaussian mixture with K components, as explained above. The number of Gaussian mixture components (K) can theoretically vary between each predicted distribution, but in this example, for simplicity, the inventors set it to be equal and adjusted K via cross-validation. The final output of the model is the full distribution parameter set as described above. The ESM2 Transformer model described in Rives et al. (2021), PNAS 118(15)e2016239118 (available at https: / / github.com / facebookresearch / esm) was used to initialize embedding models for both peptide and MHC sequences. Two copies of the model were included to encode the peptide sequence and the MHC pseudo-sequence, respectively. The ESM2 embeddings were averaged at residue sites and appended with a small MLP head with a single (logit) output. The resulting classifier model was trained using data from Chu et al. (2022) and a binary cross-entropy / log loss function to predict peptide-MHC binding. The binary classification task was designed to provide an understanding of efficient binding interfaces for multiple peptides and HLA against a model pre-trained on a general protein-based model. The classification head was then removed and replaced with a new regressor MLP. Supervising this regression task requires 3D structures, and only a few hundred such structures are available in the PDB. Therefore, the inventors defined each training instance as a single HLA residue and a single peptide residue. Given a residue pair, the model extracts slices of the corresponding residues from the ESM embeddings, concatenates them, and passes the resulting tensor through a regressor MLP. Training is performed via maximum likelihood estimation (MLE) using the negative logarithm of the distribution as the loss function. As explained above, w is an adjusted hyperparameter that balances distance loss and dihedral angle loss, and the subscript θ denotes the neural network (MLP) weight parameter. The dataset used for training the regressor MLP includes structures with missing distance or angle labels. The inventors considered the missing labels during training, such that when the D of HLA residue i and peptide residue j... ij When the distance tag is missing, or when the φ of peptide residue j is... j and ψ j When both labels are missing, the corresponding model prediction is not included in the loss calculation. However, when φ is observed... j and ψ j When one of the tags is available, the residue is not discarded, and the joint distribution in the predicted distribution of the dihedral angle is decomposed into p(φ) based on the availability of the tag. j | ψ j ) p(ψ j ) or p(ψ j | φ j ) p(φ j Because the marginal term p(φ) j Even when only φ is observed j Evaluation is also allowed during labeling, and for ψj Similarly, the model can still learn from residues lacking a single angular label. During training, residues with both φ and ψ labels are randomly assigned to evaluate either condition p(φj | ψj) or p(ψj | φj), and residues with only one angular label are assigned to evaluate marginal terms. The optimal hyperparameter configuration is determined via 5-fold cross-validation. The inventors plotted 1000 predicted samples of the structure descriptor. These samples were then merged in 5-fold cross-validation to evaluate the performance of the method.

[0202] Molecular dynamics simulations: MD simulations were used to evaluate the model results. MD simulations were initiated from the PDB structure dataset compiled above. Simulations were performed on the bound peptide-HLA complex. HLA was truncated (residues 1-180) to improve simulation speed. Hydrogen atoms were stripped, and the side chain pKa was predicted using PROPKA at pH 7.4, with protonated states assigned using PDB2PQR and an Amber FF19SB force field. The protein was solvated in an octahedral box of OPC water, extending 10 Å from the protein and containing neutralized NaCl ions. These structures were relaxed by minimizing the conjugate gradient energy over 2000 steps, followed by heating to 300 K in an NPT ensemble at 0.5 ns using a force of 10 kcal mol⁻¹ Å. -2 The protein was constrained by a harmonic confinement potential. An additional 2000 steps of conjugate gradient energy minimization were performed without any constraints. After minimization, the system was equilibrated using NPT, where the pressure was maintained at 1 atm and the thermostat was at 300 K for 0.5 ns, applying 10 kcal mol⁻¹ Å to the protein. -2 The system is constrained. Then the system operates at 1 kcal mol⁻¹ Å. -2The constraint constants were balanced for 1 ns, and a final unconstrained equilibrium was achieved at 300 K. All constraints were removed during the production phase. Production simulations were performed using NPT conditions and run with Amber's hydrogen mass redistribution option, allowing a time step of 4 fs. Langevin dynamics were used to maintain the temperature at 300 K and the collision frequency at 3 ps⁻¹. Periodic boundary conditions were applied in all directions, electrostatic interactions were calculated using the Particle Mesh Ewald (PME) method with a real space truncation of 10 Å, and hydrogen lengths were constrained using the SHAKE algorithm. The simulations were run for 300 ns, repeated twice, for a total of 600 ns. The inventors analyzed the MD simulations to gain insights into the convergence of the simulations and to understand the set of conformations sampled in the MD dataset. Unless otherwise stated, the analysis was performed using cpptraj. Root mean square deviation (RMSD) values ​​were calculated for the main chain atoms (CA, C, O, N) using the 'rms' CPPTRAJ command after truncating the HLA superposition. The root mean square fluctuation (RMSF) of the peptide is calculated using the atomicfluct command in CPPTRAJ. The dihedral angle of the peptide is calculated using the 'dihedral' command.

[0203] Peptide backbone sampling was performed using FPD: The set of conformational isomers was sampled using a refinement procedure within the FPD scheme of the Rosetta software suite. The refinement procedure used a coarse-grained model of interacting chaperones. Here, the peptide backbone and its side chains were iteratively optimized, binding the peptide to a rigid body placement relative to HLA. The inventors modified the sampling temperature to 0.2 (the default is 0.8), which is used in the Metropolis criterion, to allow sampling of diverse peptide backbones during simulation. The coarse-grained model used homology modeling based on structural templates selected from: (i) sequence similarity between the target and template peptide sequences calculated using BLOSUM62, or (ii) predicted distance and dihedral constraints. For the latter, structural templates were selected that matched the target peptide length in the PDB and had no missing density. For each template, the mean Gaussian log-likelihood of the distance and dihedral was calculated as described above. Using the target peptide sequence, HLA alleles, and template structures of the complex as input, Rosetta was used for homology modeling of the target sequence of interest. The target and template sequences were aligned using ClustalOmega. The alignment results, along with the template coordinates, were then used by Rosetta's threading scheme to generate a homology model. The model underwent a further restricted refinement stage, where only peptide residues and HLA residues in contact with the peptide (within 3.5 °A) were refined in the Rosetta force field.

[0204] Structural clustering of the peptide backbone: To cluster the peptide backbone, the pairwise cosine difference matrix of the dihedral angles of the peptides is calculated as described above. Agglomerative clustering is then used to cluster the backbone. Cluster centers are then identified, and the samples closest to these cluster centers are selected as centroids for downstream tasks.

[0205] The predicted structural model is evaluated based on, for example, the following: (i) calculating the mean dihedral distance of the modeled peptide backbone relative to the "grounded truth value" based on crystal structure data of the pMHC complex; (ii) calculating the root mean square deviation (RMSD) of the dihedral distance relative to the "grounded truth value" to check the appropriate placement of the peptide relative to MHC in the pMHC complex; and (iii) determining the contact ratio of contact recovery relative to the crystal structure in the model. Dihedral distance (or dihedral angle distance) is a measure of the distance between two dihedral angles. For two dihedral angles θ1 and θ2 of the same type at the same residue location in two different structures, the distance between these two dihedral angles is defined as: (See, for example, North et al. (2011), “A New Clustering of Antibody CDR LoopConformations”, J. Mol Biol. 406(2):228-256). The D-score is used to capture the alignment-free distance between two peptide backbones. The D-score is calculated for the φ, ψ, and ω dihedral angles of L-1 residues of peptide length L. Here, A and B are peptides of length L. The peptide dihedral angles of the predicted complex and the MD complex represent samples drawn from two distributions. The D-score described above quantifies the differences between these distributions by treating angles φ, ψ, and ω as independent—calculating the angular distance for each position and summing it in a simple manner. However, position-wise metrics such as the D-score do not consider the correlation of the three dihedral angles across residue positions. The root mean square deviation (RMSD) between two structures (such as, for example, the predicted structure and the “grounded truth” structure) is calculated as... , where n is the number of isovalent atomic pairs in the two alignment / superposition structures, and di is the distance between the two atoms in the i-th pair (usually in Å, typically Euclidean distance). RMSD can be calculated for any type and / or subset of atoms in a protein or protein-peptide complex. Unless otherwise stated, the RMSD values ​​provided herein refer to all main chain heavy atoms (C, N, C6) of both the peptide and protein in the peptide-protein complex. αThe RMSD calculated by (and O) for this peptide-protein complex was compared between two structures. Side chain contact recovery was evaluated between the predicted model and the MD model or crystal structure. The number of HLA-peptide contacts in the reference structure (MD model or crystal structure) was calculated as follows: First, residues on the HLA whose Cα is within 15 Å of the peptide residue were identified (or neighboring HLA residues were identified to narrow the search space for subsequent steps). Then, any atom-to-atom contacts from neighboring HLA and peptide residues were captured if they were within a threshold of 3.1 Å. The proportion of contacts recovered by the model was then estimated by counting the overlapping contacts between the reference structure and the model of interest.

[0206] Figure 11A Non-limiting examples are provided for a benchmark set of pMHC complexes containing peptides of varying lengths (8, 9, 10, 11, and 13 amino acid residues) (plotted on the x-axis – identified by the PDB identifier of the corresponding structure), with data (plotted on the y-axis) showing the predicted root mean square deviation (RMSD) of the Φ and Ψ dihedral angles of the peptide-protein complexes relative to the corresponding “grounded truth” crystal structures, where the Φ and Ψ dihedral angles were treated independently during training and inference. The results are shown as the distribution of RMSD values ​​for multiple conformational isomer structures predicted as described herein. In this early example, approximately 25% of the predicted structures exhibit RMSDs greater than 2.5 Å. Figure 11B A non-restrictive comparative example is provided for a benchmark set of pMHC complexes containing peptides of different lengths (8, 9, 10, 11, and 13 amino acid residues) (plotted on the x-axis), showing the root mean square deviation (RMSD) of the predicted Φ and Ψ dihedral angles of the peptide-protein complexes compared to the corresponding “grounded truth” crystal structures (plotted on the y-axis), where the Φ and Ψ dihedral angles were treated independently or jointly for a given amino acid residue during training and inference. The results are shown as the distribution of RMSD values ​​for multiple conformational isomer structures predicted as described herein. Joint treatment of the dihedral angles resulted in more accurate (reduced RMSD) predictions. Figure 11CA non-restrictive comparative example is provided for a benchmark set of pMHC complexes containing peptides of varying lengths (plotted on the x-axis), showing the root mean square deviation (RMSD) of the predicted dihedral angles Φ and Ψ of the peptide-protein complexes compared to the corresponding “grounded true” crystal structures (plotted on the y-axis). One dataset (red single dots) was generated using a previous protein structure prediction model (AlphaFold – FineTune (AF-FT)), and the other dataset (green dot distribution) was generated using the method disclosed herein (joint prediction with dihedral angles). The results are shown as the distribution of RMSD values ​​for multiple conformational isomer structures predicted as described herein. These preliminary results indicate that AF-FT accurately predicts crystal structures but fails to reflect the dynamics of pMHC complexes in solution. Figure 11D Non-limiting examples of preliminary data for predicting the mean dihedral distance are provided, plotted on the y-axis of a set of benchmark pMHC complexes (plotted on the x-axis) containing peptides of varying lengths (8, 9, 10, 11, and 13 amino acid residues). These results are consistent with the expectation that longer peptides have greater flexibility. Figure 11E Non-limiting examples of preliminary data for recovering crystal structure contacts (i.e., the proportion of amino acid residues in the crystal structure with a distance of less than 3.1 Å that also have a distance of less than 3.1 Å in the predicted structure) of a baseline pMHC complex set (containing peptides of lengths of 8, 9, 10, 11, and 13 amino acid residues) are provided, where the predicted structures are obtained using the disclosed structure prediction method. The distribution of the recovered contact proportions on the predicted conformational isomer structures (y-axis) is plotted against the baseline pMHC complex (x-axis). The data show a high correlation between the predicted contacts and the crystal structure contacts.

[0207] To evaluate the structure descriptors predicted by the method in this example, the inventors partitioned the training and test sets based on sequence and structural similarity. For the sequence-based partitioning, the training and test sets consisted of 80% and 20% of the total 3D structure data from the PDB, respectively, with peptide lengths up to and including 11 AA. The training set was further recombined into 5-fold cross-validation sets, consisting of approximately 80% and 20% of the 80% structure data, respectively. The test and validation sets consisted of data point bins covering all alleles and peptide lengths with more than one co-crystal structure. For peptides longer than 11 AA, the inventors trained the model based on data from the PDB and a structure model generated using Alphafold-Finetune. The selection of the validation set followed a similar strategy as described above when training the model using only structure data from the PDB. To generate structure-based data partitioning, the inventors binned the structures by peptide length (up to 10 AA). For each peptide length, they clustered the peptide backbone using a D-score radius of 5.9. All singleton clusters were moved to the test set. From the remaining clusters, each randomly selected member was placed into a randomly selected validation set until a count of 60 was reached in the validation set. Sequence-based partitioning probes the model's ability to reproduce structural descriptors of peptides with sequences not seen in the training set. Structure-based partitioning stress-tests the model's ability to generalize to structural clusters not presented in the training set, particularly for longer peptides (11-13 amino acids). These test sets consist of PDB structures that are diverse in both sequence and structural space and different from those in the training set, ensuring reliable model evaluation. Figure 12 shows exemplary results demonstrating the method's ability to recover dihedral angles and distances. Figure 12A Box plots are shown to highlight the average predicted distances (sampled from a aggregated Gaussian distance distribution) of peptide positions grouped by peptide length. The box plots show a distance distribution constructed from the minimum (the minimum of 34 distances from peptide residues to HLA residues) predicted average distance, which is the average of each peptide position for each test instance across all modeled samples (200 FPD models in a 5-fold distribution). The X-axis in Figures A and B represents peptide positions. Figure 12AAs shown, the method produced accurate predictions of distances to shorter peptides in both data partitions. The model is able to distinguish between anchored and non-anchored residue positions. The predicted distances highlight that anchored residues generally have a more compact distribution and shorter distances—for example, positions 1 and 8 for 8-mers, positions 1, 2, and 9 for 9-mers, and positions 1, 2, 9, and 10 for 10-mers (Figure 12A; these distributions show the average minimum distances from peptide residues to any HLA residue in the FPD sampled structures from the test set). Non-anchored residues have larger distances. Next, to analyze the coverage of the estimated dihedral density, the inventors overlaid the native dihedrals in the test set with the predicted joint density, where the y-axis represents ψ and the x-axis represents φ (or a Ramachandran plot). Figure 12B The Ramachandran plot shows the dihedral recovery for shorter peptides of length 8, 9, or 10 (shown as "short") or longer peptides of length 11, 12, or 13 (shown as "long"). The predicted dihedral density is shown by a light gray outline (covering the allowed region of 98% of residues) or a dark gray outline (covering the preferred region of 98% of residues). Scatter dots represent the true dihedrals in the test set. The color and shape of these dots depend on whether they fall within the allowed region (blue circles) or outside the allowed region (red crosses). Figure 12B The results show that for all instances in the test set (including OOD cases), there is a consistent and complete overlap between the predicted distribution and the natural dihedral angle, with at least 99.95% of the predicted samples falling within the allowable region. To further highlight that the method has learned the correct descriptor, Figure 12CA representative example, 1I1Y (9-mer), is shown in Figure 12C. The position-by-position Ramachandran plot of the representative 9-mer example 1I1Y is shown. The allowed region (light green) covers 99.95% of the dihedral angles, while the preferred region (dark green) covers 98%. Peptide positions are highlighted using black circles on the bar chart shown in wheat-colored. These plots highlight the results of two data partitions marked with orange (sequence-based) or green (structure-based) circles. Figure 12C shows the predicted dihedral distribution tightly clustered along anchor residues (residue positions 2 and 8) and more dispersed along intermediate residues (residue positions 5 and 6), reflecting the typical situation for 9-mer peptides in crystal structures. This dispersion along intermediate residues allows for a wider sampling space and thus helps capture the plasticity of the peptide. The more dispersed distance distribution compared to shorter peptides is as expected, due to the greater flexibility of longer peptides (Figure 12A). This model can distinguish between anchored and non-anchored positions, which are typically located in the N-terminal and C-terminal regions of peptides. Here, multiple stable interactions with HLA grooves were found along amino acids at positions 1–4 (Fig. 12A), which is reflected in the crystal structure. For all long peptide tests, the predicted dihedral angles overlapped with the natural dihedral angles (Fig. 12B) and spanned the allowable range of the ψ angle. This degeneracy along the predicted dihedral angles and distances allows for sampling of diverse peptide backbones.

[0208] To evaluate the performance of this method, the inventors sampled a set of structures from two datasets divided into two sets, guided by ML constraints, using FPD. The set of conformational isomers was obtained using one or more templates and refined in FPD. Templates for a given target complex were selected by: (1) using sequence similarity (or sequence homology), (2) using structural similarity estimation relative to the cocrystal structure in the PDB based on the predicted structural descriptor of the target complex, and / or (3) clustering and locating cluster centers for available peptide backbones of the target length from the PDB or MD samples. Generally, for shorter peptides (up to 10 amino acids), the inventors found in this example dataset that selecting a single template based on sequence or structural similarity performed very well. Not wanting to be bound by theory, the inventors believe this is because peptides in HLA grooves are anchored and stretched, allowing some plasticity along intermediate residues. For longer peptides (i.e., 11 amino acids and longer), selecting multiple different templates helps allow sampling of diverse peptide backbones. This is because these peptides are expected to be highly flexible, and therefore using multiple starting points can help better capture their flexibility. As a baseline, the inventors sampled a set of structures from a benchmark set, starting with homologs selected based on sequence similarity and refining them without any guidance (i.e., without using constraints predicted by the ML methods described herein). The inventors demonstrated the success of this method by its ability to recover i) the native peptide backbone and ii) overlapping and non-overlapping peptide backbones sampled from MD simulations. Since the baseline uses homologs as templates, better recovery of the native cocrystal structure is generally expected because a small number of mutant peptides beyond the edit distance are available and retain the desired peptide backbone. The inventors used D-scores and distribution distances as metrics to evaluate the model from this method. A peptide backbone is considered recovered if at least one FPD sample has a D-score of 10 or higher from any of the 300 MD samples. The recovery rates for 8-mer, 9-mer, and 10-mer from sequence-based partitioning were 99.9%, 83.9%, and 69%, respectively; the recovery rates for 8-mer, 9-mer, and 10-mer from structure-based partitioning were 100%, 69.7%, and 43.8%, respectively; and the recovery rates for all 11-mer, 12-mer, and 13-mer in the PDB were 71.7%, 47.5%, and 82.6%, respectively. Here, the template selection strategy was: 8-mer and 9-mer based on constraints, 10-mer based on sequence, and 11-mer, 12-mer, and 13-mer as multiple templates, followed by constraint refinement using predictions from a machine learning model.

[0209] Figure 13Non-limiting example data are presented to demonstrate that the structural models sampled by this method are diverse across the benchmark set for all peptide lengths and overlap with the MD set. Figure 13 is a plot showing the number of FPD models overlapping with MD samples for all PDBIDs (x-axis) in the benchmark set (y-axis; using a D-score cutoff of 10.0), for peptide lengths 8 (red), 9 (magenta), 10 (blue), 11 (orange), 12 (red), and 13 (brown) (from left to right, indicated by the color bars below the x-axis). The yellow and gray bars (the left and right bars in each pair for the benchmark complex) represent constrained refinement and unconstrained refinement, respectively. The data show that the constrained refinement strategy recovers the MD backbone better than the baseline strategy for shorter peptides in at least 33% and 50% of the benchmark cases. Similarly, for longer peptides, constrained refinement recovers more of the MD backbone, particularly in the case of 19%, while the baseline is better or comparable in the cases of 1% or 80%. Furthermore, FPD samples that do not have backbone overlap with MD are not necessarily incorrect; if they have a low Rosetta energy (indicating overall complex stability) and a binding score, they may be legitimate backbones. Total energy and binding energy can also be used to filter out samples that may be problematic and have poor total energy and binding energy readings.

[0210] Figure 14A -B provides non-limiting example data demonstrating that the set of conformational isomers sampled by the methods of this disclosure approximates the baseline MD distribution. Figure 14A-B shows the distribution of evaluation metrics for the FPD model relative to native crystal structures for all peptide lengths using constrained and unconstrained refined sampling. Along the y-axis, the figure shows: (top) RMSD (in Å) of peptide backbone heavy atoms (N, Cα, C, and O); (middle) D-score values ​​(0 indicates complete recovery of the peptide backbone); (bottom) recovery of side chain contacts of the FPD model relative to the baseline PDBID (along the x-axis) of the native crystal structure (0 indicates no recovery, while 1 indicates perfect side chain placement relative to native). Sequence-based partitioning data are shown in Figure 14A, and structure-based partitioning data are shown in Figure 14B, applicable to test PDBs for 8–10 AA and all 11–13 AA. Peptide lengths 8 (red), 9 (magenta), 10 (blue), 11 (orange), 12 (red), and 13 (brown) are indicated by color bars below the x-axis (in the order above from left to right). As discussed earlier, the baseline method is expected to sample natural structures better; however, the data shows that constraints during refinement guide the model to sample more closely to natural structures. Specifically, Figure 14A shows the results of the sequence-based partitioning test set, highlighting 25% of 8-mers (4F7T), 30% of 9-mers (1QRN, 2C7U, 2VLL, 2X4O, 3BVN, 3D18, 3I6G, 3I6L, 3KPM, 4MNQ, 4N8V, 4O2E, 4QRS, 5ISZ, 5MER, 5TEZ, 5VVP, 5WMO, 6RPA, 6Z9W, 7KGP, 7KGS, 7LGD, 7NME, 7R80, 7ZUC), and 25% of 10-mers (1HHH, 5C09, 6MPP, 7S8Q, 7S8S, 8DVG), as highlighted by peptide backbone heavy atom RMSD, D-score, or side chain contact recovery (Figure 14A). Furthermore, similar trends can be observed in the structure-based partitioning test set for 50% of 8-mers (3BWA), 30% of 9-mers (1XR9, 2X4S, 3D18, 7L1C), and 48% of 10-mers (2CLR, 3GIV, 3MRN, 3UTT, 4F7M, 5VZ5, 6D29, 6MPP, 7JYU, 7S8S) (Figure 14B).Finally, among the longer peptides in the test set consisting of all available structures in the PDB, 43% of the 11-mers (2FZ3, 2HJK, 3MV7, 3MV8, 4PRB, 4PRD, 4PRE, 4PRI, 5D9S, 5WWI, 5WWU, 6D2R, 6VPZ), 80% of the 12-mers (3BW9, 4JQX, 5DDH, 5T6Y), and 14% of the 13-mers (3VFS, 6AVF) had sampled conformations closer to the native structures than the unconstrained refinement setting (Figure 14A).

[0211] Figure 16A -B provides non-restrictive examples of preliminary control data for pMHC template structure selection and refinement, respectively. Figure 16A Provides the C, N, and C heavy atoms of the peptide backbone in the selected starting template structure. α A plot of the median simulated RMSD of the pairwise distance between O (for peptides of lengths of 8, 9, 10, 11, and 13 amino acid residues) and the corresponding RMSD of the pairwise distance obtained from crystal structure data. Figure 16B Provides the C, N, C heavy atoms of the peptide backbone in the generated pMHC conformational isomer set. α A plot of the median simulated RMSD of the pairwise distances between the pMHC and O (for peptides of lengths of 8, 9, 10, 11, and 13 amino acid residues) versus the corresponding RMSDs of the pairwise distances obtained from crystal structure data is presented, illustrating the dynamic properties of the pMHC structure in solution.

[0212] Figure 17A -C provides peptide conformation sampling (Φ (x-axis) and Ψ (y-axis) dihedral plots) in the pMHC complex. Figure 17A Molecular dynamics data (300 ns simulation) of the pMHC complex (PDBID: 2BCK; 9-mer peptide) and graphical representation of the pMHC complex (PDBID: 2BCK; 9-mer peptide). Figure 17B The graph shows the variation of the peptide backbone fluctuation (root mean square fluctuation (RMSF)) of the pMHC complex with residue numbering (residues 1 to 181 correspond to MHC molecules, and subsequent residues correspond to peptides). Figure 17C Non-restrictive examples of ). Figure 17CA graph showing the minimum and maximum RMSF values ​​obtained for hundreds of 9-mer peptide sets as a function of peptide residue numbering is provided. The data demonstrate the differences in peptide flexibility across the pMHC complex as a function of peptide residue position. Root mean square fluctuation (RMSF) is a measure of how much an individual atom or group of atoms in a molecule fluctuates from its average position over time, essentially indicating the flexibility of a specific region within the structure. This is most commonly used in molecular dynamics simulations to analyze protein dynamics, where regions with high RMSF values ​​are considered more flexible.

[0213] Figure 18A -C provides peptide conformation sampling (Φ (x-axis) and Ψ (y-axis) dihedral plots) in the pMHC complex. Figure 18A Molecular dynamics data (300 ns simulation) of the pMHC complex (PDBID: 3VFT; 13-mer peptide) and graphical representation of the pMHC complex (PDBID: 3VFT; 13-mer peptide). Figure 18B The graph shows the variation of the peptide backbone fluctuation (root mean square fluctuation (RMSF)) of the pMHC complex with residue numbering (residues 1 to 181 correspond to MHC molecules, and subsequent residues correspond to peptides). Figure 18C Non-restrictive examples of ). Figure 18C A graph showing the RMSF (Residual Mobility Spectroscopy) of 11 different 13-mer peptides as a function of peptide residue numbering is provided. The data demonstrate the differences in peptide flexibility within the pMHC complex as a function of peptide residue position.

[0214] Figure 15A -B illustrates a non-limiting example demonstrating that using predictive constraints to guide sampling of peptide conformations, as described herein, can elucidate states that determine TCR specificity. To explore the ability of the methods described herein to sample relevant protein conformations, the inventors used the HLA-A03:01 molecule complexed with a mutant PIK3CA peptide as an example. In a recent study (Ma, J., et al., Nature Communications 16(1), 849 (2025)), the authors discussed a significant conformational change in the side chain of position 6 tryptophan (pTrp6) upon binding. In the unbound state, pTrp6 abuts against the HLA-A3 α1 helix ( Figure 15A (pink image), but during bonding, the pTrp6 side chain flips around axis a and abuts against the α2 helix ( Figure 15A(Purple figure). A significant energy barrier exists around the main chain inversion, making it difficult for most models to overcome. Dynamic allosteric changes are assumed to be common in TCR recognition of peptides / HLA, highlighting the importance of accurately predicting different conformational samples when predicting cross-reactivity and specificity. Here, the inventors explored the ability of the method described herein to sample the bound and unbound conformations of PIK3CA peptides. They performed 300 ns simulations of HLA-A03:01 complexed with the mutant PIK3CA peptide (PDBID:7L1D) and HLA-A03:01 complexed with the mutant PIK3CA peptide (PDBID:7L1D) without TCR. Conventional MD simulations cannot overcome the energy barrier associated with conformational changes such as side chain inversion (…). Figure 15B These limitations underscore the importance of incorporating constraints during refinement to achieve a more comprehensive sampling of the conformational landscape. When constrained guidance is used, the methods described in this paper effectively model the dynamic properties of peptide-HLA complexes, capturing both bound and unbound states. Figure 15A shows the structural representation of HLA-A03:01 in different states. The pink plot in Figure 15A (PDB ID: 7L1C) represents unbound HLA-A03:01 complexed with the mutant PIK3CA peptide. The purple plot in Figure 15B (PDB ID: 7L1D) depicts the TCR of HLA-A03:01 complexed with the mutant PIK3CA peptide. The gray plot in Figure 15A shows the modeled HLA-A03:01 complexed with the mutant PIK3CA peptide, exhibiting multiple conformations, highlighting the model's ability to sample both bound and unbound conformations observed in the crystal structure. The pink plot in Figure 15B (PDB ID: 7L1C) represents unbound HLA-A03:01 complexed with the mutant PIK3CA peptide. The purple plot in Figure 15B (PDB ID: 7L1D) depicts the TCR of HLA-A03:01 complexed with the mutant PIK3CA peptide. The gray plot in Figure 15B shows the modeled HLA-A03:01 complexed with the mutant PIK3CA peptide, presenting the highest RMSD state from the MD 7L1C simulation. This highlights that the MD simulation cannot overcome the energy barrier associated with side-chain flipping.

[0215] The intrinsic plasticity of peptides displayed on HLA molecules drives TCR specificity and cross-reactivity. Nevertheless, modeling the conformational flexibility of such protein-ligand complexes without employing time-consuming MD simulations has been particularly challenging. Furthermore, the newly developed time-saving methods focus on monomers or small molecules. In these examples, the inventors combined the advantages of machine learning and molecular modeling tools (e.g., Rosetta (FPD)) and sampled peptide conformations of 8–13 amino acids presented by various HLA alleles. They showed that they could sample peptide backbones that overlapped with the backbones of MD simulations from most examples in a blinded test set. Various metrics, including alignment-free D-score measures, were used to highlight the overlap between MD and the predicted set of conformational isomers. The inventors hypothesized that they should be able to model peptide plasticity in pHLA under constraint guidance, which could limit a large sampling space. Therefore, they employed a machine learning model trained on pHLA binding sequences and complexes to predict structural descriptors, particularly the joint distribution of distances and dihedrals used as sampling constraints. Here, the ML model uses an ESM2 sequence embedding shared between the classifier (learning amino acid associations between peptides and HLA) and the regressor (predicting structural constraints). Compared to the baseline model (unconstrained), they show that the constraint-guided model is better at recovering MD samples and providing additional reasonably diverse peptide backbones regarding peptide plasticity within HLA grooves.

[0216] Example computer system Figure 19 Non-limiting examples of block diagrams of computer systems according to some embodiments are provided. Computer system 1900 may be as described above. Figure 1 An example of one implementation of the computing platform 102 described herein.

[0217] Computer System 1900 can be a host computer connected to a network. Computer System 1900 can be a client computer or a server. For example... Figure 19 As shown, computer system 1900 can be any suitable type of microprocessor-based device, such as a personal computer, workstation, server, or handheld computing device (portable electronic device), such as a telephone or tablet. The device may include, for example, one or more of processor 1910, input device 1920, output device 1930, storage 1940, and communication device 1960. Input device 1920 and output device 1930 may generally correspond to devices described elsewhere herein, and they may be connectable or integrated with a computer.

[0218] Input device 1920 can be any suitable device that provides input, such as a touchscreen, keyboard or keypad, mouse, or voice recognition device. Output device 1930 can be any suitable device that provides output, such as a touchscreen, haptic feedback device, or speaker.

[0219] Storage 1940 can be any suitable device providing storage, such as electrical memory, magnetic memory, or optical memory including RAM, cache, hard disk drive, or removable storage disk. Communication device 1960 can include any suitable device capable of transmitting and receiving signals over a network, such as a network interface chip or device. Computer components can be connected in any suitable manner, such as via a physical bus 1970 or wireless connection.

[0220] The software 1950, which can be stored in memory / storage 1940 and executed by processor 1910, may include, for example, programming that embodies the functionality of this disclosure (e.g., as embodied in the methods described above).

[0221] Software 1950 may also be stored and / or transferred in any non-transitory computer-readable storage medium for use by or in connection with an instruction execution system, device, or apparatus (such as those described above), from which the instruction execution system, device, or apparatus may retrieve and execute instructions associated with the software. In the context of this disclosure, the computer-readable storage medium may be any medium (such as storage 1940) that may contain or store programming for use by or in connection with an instruction execution system, device, or apparatus.

[0222] Software 1950 can also be propagated within any transmission medium for use by or in conjunction with an instruction execution system, device, or apparatus (such as those described above), from which the instruction execution system, device, or apparatus may obtain and execute instructions associated with the software. In the context of this disclosure, the transmission medium can be any medium capable of communicating, propagating, or transmitting programming for use by or in conjunction with an instruction execution system, device, or apparatus. Transmission readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, or infrared wired or wireless propagation media.

[0223] The computer system 1900 can be connected to a network, which can be any suitable type of interconnected communication system. The network can implement any suitable communication protocol and can be protected by any suitable security protocol. The network can include any suitable arrangement of network links that can enable the transmission and reception of network signals, such as wireless network connections, T1 or T3 lines, cable networks, DSL, or telephone lines.

[0224] Computer system 1900 can implement any operating system suitable for operation on a network. Software 1950 can be written in any suitable programming language, such as C, C++, Java, or Python. In various embodiments, application software embodying the functionality of this disclosure can be deployed in different configurations, such as, for example, in a client / server setup, or as a web-based application or web service via a web browser.

[0225] Figure 20 An exemplary artificial intelligence (AI) architecture 2002 according to the disclosed embodiments is shown (which can serve as a reference above regarding...). Figure 19 Figure 2000 (included in the discussion of one or more computing devices 1900) illustrates an artificial intelligence architecture that can be used to determine one or more predicted peptide-protein binding interactions and / or generate multiple compatible structures of peptide-protein complexes. In some embodiments, the AI ​​architecture 2002 may be implemented using, for example, one or more processing devices, which may include hardware (e.g., general-purpose processors, graphics processing units (GPUs), application-specific integrated circuits (ASICs), system-on-a-chip (SoCs), microcontrollers, field-programmable gate arrays (FPGAs), central processing units (CPUs), application processors (APs), vision processing units (VPUs), neural processing units (NPUs), neural decision processors (NDPs), deep learning processors (DLPs), tensor processing units (TPUs), neuromorphic processing units (NPUs), and / or other processing devices that may be adapted to process various molecular data and make one or more decisions thereon), software (e.g., instructions that run / execute on one or more processing devices), firmware (e.g., microcode), or some combination thereof.

[0226] In some embodiments, such as Figure 20The depicted AI architecture 2002 may include machine learning (ML) algorithms and functions 2004, natural language processing (NLP) algorithms and functions 2006, expert systems 2008, computer-based vision algorithms and functions 2010, speech recognition algorithms and functions 2012, planning algorithms and functions 2014, and robotics algorithms and functions 2016. In some embodiments, the ML algorithms and functions 2004 may include any statistically based algorithms applicable to finding patterns in large amounts of data (e.g., “big data,” such as genomic data, proteomic data, metabolomic data, metagenomic data, transcriptomic data, or other omics data). For example, in some embodiments, the ML algorithms and functions 2004 may include deep learning algorithms 2018, supervised learning algorithms 2020, and unsupervised learning algorithms 2022.

[0227] In some embodiments, the deep learning algorithm 2018 may include any artificial neural network (ANN) that can be used to learn deep representations and abstractions from large amounts of data. For example, the deep learning algorithm 2018 may include ANNs such as perceptrons, multilayer perceptrons (MLPs), autoencoders (AEs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTMs), gated recurrent units (GRUs), restricted Boltzmann machines (RBMs), deep belief networks (DBNs), bidirectional recurrent deep neural networks (BRDNNs), generative adversarial networks (GANs) and deep Q-networks, neural autoregressive distribution estimation (NADEs), adversarial networks (ANs), attention models (AMs), spiking neural networks (SNNs), deep reinforcement learning, etc.

[0228] In some embodiments, the supervised learning algorithm 2020 may include any algorithm that can be used to apply, for example, previously learned content to new data using labeled examples to predict future events. For example, starting with the analysis of a known training dataset, the supervised learning algorithm 2020 may generate an inference function to predict output values. The supervised learning algorithm 2020 may also compare its output with correct and expected outputs and identify errors in order to modify the supervised learning algorithm 2020 accordingly. On the other hand, the unsupervised learning algorithm 2022 may include, for example, any algorithm that can be applied when the data used to train the unsupervised learning algorithm 2022 is neither classified nor labeled. For example, the unsupervised learning algorithm 2022 may study and analyze how the system infers a function describing the hidden structure from unlabeled data.

[0229] In some embodiments, NLP algorithms and functions 2006 may include any algorithms or functions suitable for automatically processing natural language, such as speech and / or text. For example, NLP algorithms and functions 2006 may include content extraction algorithms or functions 2024, classification algorithms or functions 2026, machine translation algorithms or functions 2028, question answering (QA) algorithms or functions 2030, and text generation algorithms or functions 2032. In some embodiments, content extraction algorithms or functions 2024 may include means for extracting text or images from electronic documents (e.g., web pages, text editor documents, etc.) for use, for example, in other applications.

[0230] In some embodiments, the classification algorithm or function 2026 may include any algorithm that can learn from the data input of a supervised learning model and make new observations or classifications based thereon, utilizing a supervised learning model (e.g., logistic regression, Naive Bayes, stochastic gradient descent (SGD), k-nearest neighbors, decision trees, random forests, support vector machines (SVMs), etc.). The machine translation algorithm or function 2028 may include any algorithm or function that can be adapted to automatically convert source text in one language, for example, into text in another language. The QA algorithm or function 2030 may include any algorithm or function that can be adapted to automatically answer questions posed by humans, for example, in natural language, such as algorithms or functions performed by a voice-controlled personal assistant device. The text generation algorithm or function 2032 may include any algorithm or function that can be adapted to automatically generate natural language text.

[0231] In some embodiments, expert system 2008 may include any algorithms or functions applicable to simulating the judgment and behavior of a person or organization with expertise and experience in a specific domain (e.g., stock trading, medicine, sports statistics, etc.). Computer-based vision algorithms and functions 2010 may include any algorithms or functions applicable to automatically extracting information from images (e.g., photographic images, video images). For example, computer-based vision algorithms and functions 2010 may include image recognition algorithm 2034 and machine vision algorithm 2036. Image recognition algorithm 2034 may include any algorithms applicable to automatically identify and / or classify objects, locations, people, etc., that may be included in one or more image frames or other display data. Machine vision algorithm 2036 may include any algorithms applicable to allowing a computer to "see," or, for example, relying on an image sensor camera with dedicated optics to acquire images for processing, analyzing, and / or measuring various data features for decision-making purposes.

[0232] In this document, “or” is inclusive rather than exclusive, unless otherwise expressly stated or the context otherwise indicates. Therefore, in this document, “A or B” means “A, B, or both”, unless otherwise expressly stated or the context otherwise indicates. Furthermore, in this document, “and” is both common and separate, unless otherwise expressly stated or the context otherwise indicates. Therefore, in this document, “A and B” means “A and B, commonly or separately”, unless otherwise expressly stated or the context otherwise indicates.

[0233] In this article, “automatic” and its derivatives mean “without human intervention”, unless otherwise explicitly indicated or indicated by the context.

[0234] The embodiments disclosed herein are merely examples, and the scope of this disclosure is not limited thereto. Embodiments of this disclosure specifically disclose, in the appended claims for methods, storage media, systems, and computer program products, any feature (e.g., method) referenced in one claim class may be claimed in another claim class (e.g., system). Dependents or references in the appended claims are chosen solely for formal reasons. However, any subject matter arising from the intentional reference to any of the foregoing claims (particularly multiple dependents) may also be claimed, thus any combination of claims and their features may be disclosed and claimed regardless of the dependents chosen in the appended claims. The subject matter that may be claimed includes not only combinations of features as listed in the appended claims, but also any other combination of features in the claims, wherein each feature referenced in the claims may be combined with any other feature or combination of features in the claims. Furthermore, any embodiments or features described or depicted herein may be claimed in individual claims and / or in any combination with any embodiments or features described or depicted herein or with any features of the appended claims.

[0235] The scope of this disclosure covers all changes, substitutions, variations, alterations, and modifications to the exemplary embodiments described or illustrated herein that will be understood by those skilled in the art. The scope of this disclosure is not limited to the exemplary embodiments described or illustrated herein. Furthermore, although this disclosure describes and illustrates corresponding embodiments herein as including specific components, elements, features, functions, operations, or steps, any of these embodiments may include any combination or arrangement of any component, element, feature, function, operation, or step described or illustrated anywhere herein that will be understood by those skilled in the art. Furthermore, references in the appended claims to an apparatus or system or a component of an apparatus or system being adapted to, arranged to, enabled to, configured to, operable to, or operated to perform a particular function cover that apparatus, system, or component, whether or not it or that particular function is activated, turned on, or unlocked, provided that the apparatus, system, or component is so adapted, arranged, enabled to, configured to, operable to, or operated. Furthermore, although this disclosure describes or illustrates specific advantages provided by certain embodiments, some embodiments may not provide, partially provide, or provide all of these advantages.

[0236] Exemplary description of the terminology As used herein, the terms “peptide,” “polypeptide,” and “protein” are used interchangeably to refer to polymers of amino acid residues. The term encompasses amino acid chains of any length, including full-length proteins with amino acid residues linked by covalent peptide bonds.

[0237] As used herein, a “mutant peptide” (or “mutated peptide”) can refer to a peptide that is not present in the normal tissue of an individual subject (e.g., in the wild-type amino acid sequence of normal tissue). A mutant peptide contains at least one mutant amino acid and may be present in diseased tissue (e.g., collected from a specific subject) but not in normal tissue (e.g., collected from a specific subject, collected from different subjects, and / or identified in a database as corresponding to normal tissue). A mutant peptide may include an epitope. An epitope is the portion of a mutant peptide that binds to an MHC molecule or TCR. Therefore, the binding between the epitope of a mutant peptide and an MHC molecule or TCR can induce an immune response (because the mutant peptide is not associated with the subject's “self”). A mutant peptide may include or be a neoantigen. Mutant peptides can originate from the following non-limiting examples: non-synonymous mutations resulting in different amino acids in a protein (e.g., point mutations); readthrough mutations in which a stop codon is modified or deleted, resulting in the translation of a longer protein with a novel tumor-specific sequence at the C-terminus; splice site mutations resulting in a unique tumor-specific protein sequence; chromosomal rearrangements producing chimeric proteins with tumor-specific sequences at the junction of two proteins (gene fusion); and / or frameshift insertions or deletions resulting in new open reading frame sequences of tumor-specific proteins. Mutant peptides may include polypeptides (characterized by polypeptide sequences) and / or may be encoded by nucleotide sequences.

[0238] As used herein, the “C-flank” of a peptide refers to one or more amino acids upstream of the C-terminus of a peptide derived from a parent protein. Optionally, the C-flank of a peptide may include one, two, three, four, five, or more amino acid residues upstream of the C-terminus of the peptide.

[0239] As used herein, the “N-flank” of a peptide refers to one or more amino acids downstream of the N-terminus of a peptide derived from a parent protein. Optionally, the N-flank of a peptide may include one, two, three, four, five, or more amino acid residues downstream of the N-terminus of the peptide.

[0240] As used herein, an epitope of a peptide can refer to the region between the C- and N-flanks of the peptide that can be recognized by the TCR. An epitope is a portion of a peptide that is recognized by the TCR on T cells and by MHC I on antigen-presenting cells. For example, an epitope can be a TCR-binding peptide, such as a TCR-binding peptide when it binds to MHC I on antigen-presenting cells.

[0241] As used in this article, a "ligand" is a peptide that is presented on the cell surface by MHC molecules in an elution experiment or found to bind to MHC in an in vitro assay.

[0242] As used in this article, "sequence" refers to an amino acid sequence that includes an ordered set of amino acid identifiers.

[0243] As used herein, a “peptide sequence” refers to a sequence of amino acids that recognize at least a portion of a peptide. In some cases, a peptide sequence includes variant-coding sequences that include variants not observed in the corresponding reference sequence.

[0244] When a peptide includes a mutant peptide, the variant-coding sequence recognizes the mutated or variant amino acid. However, when the peptide does not include a mutant or variant, the variant-coding sequence does not recognize the mutated or variant amino acid (and in this case, is identical to the reference sequence). Variant-coding sequences can be determined by collecting disease and / or tumor samples (e.g., including tumor cells) and performing sequencing analysis to identify one or more sequences corresponding to the disease and / or tumor cells in the sample. In some instances, sequencing analysis outputs an amino acid sequence. In some instances, sequencing analysis outputs a nucleic acid sequence, which can then be processed to convert codons into amino acid identifiers, thereby producing an amino acid sequence. Variant-coding sequences may include neoantigen sequences. Variant-coding sequences may, but do not necessarily, include one or more ends of the peptide (e.g., C-terminus and / or N-terminus). Variant-coding sequences may include epitopes of the peptide. Variant-coding sequences can recognize amino acids within a peptide that have one or more variants (e.g., one or more amino acid differences) relative to a corresponding reference sequence. In some cases, variant-coding sequences comprise an ordered set of amino acids. In some instances, the variant coding sequence identifies a reference peptide (e.g., by identifying a genetic reference sequence, such as by gene, start position and / or end position; or by gene, start position and / or length) and one or more point mutations relative to the reference peptide.

[0245] As used herein, a “reference sequence” can refer to a sequence (e.g., wild-type, parental sequence) that identifies at least a portion of an amino acid sequence within a non-mutant or wild-type peptide. A non-mutant or wild-type peptide may not include variants or may include fewer variants than those included in a mutant peptide. A reference sequence can include an amino acid sequence encoded by a genetic sequence within the same gene relative to a gene that includes a gene encoding a corresponding variant coding sequence. A reference sequence can include an amino acid sequence encoded by a gene sequence spanning the same start and end points within a gene, said gene sequence being associated with a gene sequence related to a corresponding variant coding sequence. Reference sequences can be identified by collecting non-disease and / or non-tumor samples from one or more subjects (who may, but need not, include subjects from whom disease samples were collected to determine variant coding sequences) and using the samples for sequencing analysis.

[0246] As used herein, a “pseudo-sequence” of an MHC molecule can refer to an ordered set of amino acids that normally contact a peptide. Therefore, a pseudo-sequence can be a subset of the amino acid sequence of a protein that is expected to contact the peptide in a peptide-protein complex. This subset can include continuous or discontinuous amino acids, but the amino acids in the pseudo-sequence are in the same order as those in the full protein sequence. A pseudo-sequence of a protein can be defined as an ordered set of amino acids in the protein sequence that are within a predetermined distance of the peptide amino acids in the peptide-protein complex structure. The structure can be determined experimentally or predicted. The amino acids included in the pseudo-sequence can also be filtered to include only polymorphic positions. Pseudo-sequences of many MHC molecules are known in the art (see, for example, Nielsen et al., NetMHCpan, a Method for Quantitative Predictions of Peptide Binding to Any HLA-A and -B Locus Protein of Known Sequence, PLoS One. 29 Aug 2007;2(8):e796).

[0247] As used herein, a “representation” or “sequence representation” of a sequence may include a set of values ​​representing or identifying amino acids in the sequence and / or a set of values ​​representing or identifying nucleic acids encoding the sequence. For example, each amino acid may be represented by a binary string and / or vector of values, which is different from each other binary string and / or vector representing each other amino acid. Sequence representations may be generated using, for example, one-hot encoding or using a block substitution matrix (BLOSUM) matrix. For example, a multidimensional (e.g., 20 or 21-dimensional) array is initialized (e.g., randomly or pseudo-randomly initialized). For each amino acid, the initialized array may include a unique vector corresponding to that amino acid. These values ​​may be fixed such that it can be assumed that such unique vectors are used to represent the corresponding amino acids.

[0248] As used herein, peptide “presentation” refers to the presentation of at least a portion of a peptide on the cell surface by binding to an MHC molecule in a specific manner. Other cells, such as nearby T cells, can then access the presented peptide.

[0249] As used herein, a “sample” may include tissue (e.g., biopsy), single-cell, multi-cell, cellular debris, or aliquots of bodily fluids. Samples may be obtained from a subject by means of, for example, but not limited to, venipuncture, excretion, ejaculation, massage, biopsy, needle aspiration, irrigation of the sample, scraping, surgical incision, intervention, and another type of sample collection method or a combination thereof.

[0250] As used herein, “subject” encompasses one or more cells, tissues, or organisms. Subjects may be human or non-human, in vivo, ex vivo, or in vitro, and may be male or female. Subjects may be mammals, such as humans. As used herein, “binding affinity” refers to the binding affinity between an amino acid (e.g., a peptide of a specific antigen) and an IPC (e.g., an MHC molecule and / or an MHC allele). Binding affinity can characterize the stability, tendency, and / or strength of binding between a peptide and an IPC molecule.

[0251] As used herein, “immunogenicity” can refer to the ability to elicit an immune response (e.g., via T cells and / or B cells). An “immunogenic” peptide can be a peptide capable of eliciting an immune response.

[0252] As used in this article, "MHC" refers to the major histocompatibility complex. Human MHC is also known as the human leukocyte antigen (HLA) complex.

[0253] The terms and expressions used are descriptive rather than restrictive, and in using such terms and expressions, no equivalent of the features shown and described or any part thereof is intended to be excluded. However, it should be recognized that various modifications are possible within the scope of the claimed invention. Therefore, it should be understood that although the claimed invention has been specifically disclosed by way of examples and optional features, those skilled in the art can employ modifications and variations of the concepts disclosed herein, and such modifications and variations are considered to be within the scope of the invention as defined by the appended claims.

[0254] Exemplary embodiments The embodiments disclosed herein may include: 1. A computer-implemented method for generating multiple compatible structures of peptide-protein complexes, the method comprising: The peptide sequence data of at least one peptide and the protein sequence data of at least one protein are input into a trained machine learning model to determine: The predicted pairwise distance data between at least one peptide amino acid residue of the at least one peptide and at least one protein amino acid residue of the at least one protein in the peptide-protein complex; and Predicted dihedral data for at least one peptide amino acid residue in at least one peptide of the peptide-protein complex; The initial structure of the peptide-protein complex is identified based on the predicted pairwise distance data and the predicted dihedral angle data; and The plurality of compatible structures of the peptide-protein complex are generated based on the initial structure, the predicted pairwise distance data, and the predicted dihedral angle data.

[0255] 2. The computer-implemented method according to Example 1 further includes predicting at least one auxiliary feature of the peptide-protein complex based on the plurality of compatible structures of the peptide-protein complex.

[0256] 3. The computer-implemented method according to Example 1 or Example 2, wherein each compatible structure of the peptide-protein complex conforms to the posterior distribution of the predicted pairwise distance data and the posterior distribution of the predicted dihedral angle data output by the trained machine learning model.

[0257] 4. The computer-implemented method according to any one of Examples 1 to 3, wherein each compatible structure of the peptide-protein complex conforms to the sequence data of the at least one peptide and the at least one protein and the prediction of at least one auxiliary feature.

[0258] 5. The computer-implemented method according to any one of Examples 2 to 4, wherein the at least one auxiliary feature includes prediction of the binding of the at least one peptide to the at least one protein.

[0259] 6. The computer-implemented method according to any one of Examples 2 to 5, wherein the at least one auxiliary feature includes the prediction of at least one interface feature of the peptide-protein complex.

[0260] 7. The computer-implemented method according to any one of Examples 1 to 6, wherein the peptide sequence data includes a peptide embedding representation of the at least one peptide, the peptide embedding representation being generated using a trained peptide embedding machine learning model.

[0261] 8. The computer-implemented method according to any one of Examples 1 to 7, wherein the protein sequence data includes a protein embedding representation of the at least one protein, the protein embedding representation being generated using a trained protein embedding machine learning model.

[0262] 9. The computer-implemented method according to Example 8, wherein the trained peptide embedding machine learning model and the trained protein embedding machine learning model are the same model.

[0263] 10. The computer-implemented method according to Example 8 or Example 9, wherein the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model includes a base model.

[0264] 11. The computer-implemented method according to any one of Examples 8 to 10, wherein the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model includes a large language model (LLM).

[0265] 12. The computer-implemented method according to any one of Examples 8 to 11, wherein the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model includes an evolutionary scale model (ESM).

[0266] 13. The computer-implemented method according to any one of Examples 8 to 12, wherein the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model are integrated with the trained machine learning model and simultaneously retrained with the trained machine learning model.

[0267] 14. The computer-implemented method according to any one of Embodiments 1 to 13, wherein the trained machine learning model comprises a trained neural network.

[0268] 15. The computer-implemented method according to Example 14, wherein the trained machine learning model includes a trained multilayer perceptron (MLP), a trained attention-based neural network, or a trained graph neural network.

[0269] 16. The method according to any one of Examples 1 to 15, wherein the trained machine learning model is trained using a training dataset comprising incomplete structural data of one or more peptide-protein complexes.

[0270] 17. The computer-implemented method according to any one of Embodiments 1 to 16, wherein the trained machine learning model is further configured to determine the uncertainty of the predicted pairwise distance data and the predicted dihedral data.

[0271] 18. The computer-implemented method according to any one of Embodiments 1 to 17, wherein the machine learning model is trained in the following manner: Receives peptide sequence data of multiple peptides; It can receive protein sequence data for a variety of proteins; Receives peptide-protein binding data for multiple peptide-protein combinations; A classifier is trained based on the peptide sequence data, the protein sequence data, and the peptide-protein binding data to predict peptide-protein binding of specific combinations of peptides and proteins; and The regressor head is attached to at least a portion of a trained classifier and trained based on pairwise distance data and dihedral angle data derived from crystal structure data of a set of exemplary peptide-protein complexes to predict the probability distribution of the pairwise distance data and dihedral angle data of the peptide-protein complexes.

[0272] 19. The computer-implemented method according to Example 18, wherein during the training of the regressor head, pairwise distance data and dihedral angle data derived from crystal structure data for each combination of individual peptide amino acid residues and individual protein amino acid residues in an exemplary peptide-protein complex are presented as separate training instances.

[0273] 20. The computer-implemented method according to Embodiment 18 or Embodiment 19, wherein the regressor is further trained based on enhanced pairwise distance data and dihedral angle data.

[0274] 21. The computer-implemented method according to Embodiment 20, wherein the enhanced pairwise distance data and dihedral angle data are obtained in the following manner: Receive peptide sequence data and protein sequence data of at least one peptide and at least one protein known to bind and form a peptide-protein complex; The peptide sequence data and protein sequence data of the at least one peptide and the at least one protein are processed using a previous version of the trained machine learning model to determine the uncertainty of the pairwise distance data and the uncertainty of the dihedral angle data of the peptide-protein complex. The determined uncertainty of the paired distance data and the determined uncertainty of the dihedral angle data are applied to the structure of an exemplary peptide-protein complex to generate multiple compatible structures of the peptide-protein complex; and Pairwise distance data and dihedral angle data of multiple compatible peptide-protein complex structures were extracted to be used as augmented training data for the regressor.

[0275] 22. The computer-implemented method according to any one of Examples 1 to 21, wherein identifying the initial structure of the peptide-protein complex based on the predicted pairwise distance data and the predicted dihedral angle data comprises: Multiple crystal structures of peptide-protein complexes were obtained from protein structure databases; and The crystal structure is selected from the plurality of crystal structures based on the similarity between the pairwise distance data and dihedral angle data determined for the crystal structure and the maximum likelihood estimates of the predicted pairwise distance data and predicted dihedral angle data determined by the trained model.

[0276] 23. The computer-implemented method according to Example 22, wherein the protein structure database includes a protein database (PDB).

[0277] 24. The computer-implemented method according to any one of Examples 1 to 23, wherein generating the plurality of compatible structures of the peptide-protein complex based on the initial structure, the predicted pairwise distance data, and the predicted dihedral angle data comprises: Input the initial structure of the peptide-protein complex into a structure modeling software package; The predicted pairwise distance data and the predicted dihedral angle data determined by the trained machine learning model for the peptide-protein complex are input into the structure modeling software package; and The structure scoring function is iteratively sampled from a plurality of possible structures of the peptide-protein complex and evaluated to identify the plurality of compatible structures of the peptide-protein complex, wherein the compatible structure of the peptide-protein complex is a structure that is (i) compatible with one or more constraints imposed by predicted pairwise distance data and predicted dihedral angle data, and (ii) optimizes the structure scoring function.

[0278] 25. The computer-implemented method according to Example 24, wherein iterative sampling from the plurality of possible structures of the peptide-protein complex is performed using a Markov chain Monte Carlo (MCMC) sampling algorithm.

[0279] 26. The computer-implemented method according to Example 25, wherein the Markov Chain Monte Carlo (MCMC) sampling algorithm includes the Metropolis-Hastington algorithm.

[0280] 27. The computer-implemented method according to Example 25, wherein the Markov Chain Monte Carlo (MCMC) sampling algorithm includes the Hamiltonian Monte Carlo algorithm.

[0281] 28. The computer-implemented method according to Example 25, wherein the Markov Chain Monte Carlo (MCMC) sampling algorithm includes the Langevin Monte Carlo algorithm.

[0282] 29. The computer-implemented method according to any one of embodiments 24 to 28, wherein the structural scoring function includes a proxy for free energy calculation.

[0283] 30. The computer-implemented method according to any one of Examples 24 to 29, wherein the structure scoring function comprises a proxy for calculating the free energy of peptide-protein binding, a proxy for calculating the free energy of at least one peptide-protein interface feature, or any combination thereof.

[0284] 31. The computer-implemented method according to any one of Examples 24 to 30, wherein the scoring function is based on one or more of the following: ab initio quantum mechanical calculation, density functional theory (DFT) calculation, semi-empirical calculation, molecular mechanical force field calculation, statistical potential calculation, neural potential calculation, or machine learning model trained on structural data.

[0285] 32. The computer-implemented method according to any one of embodiments 24 to 31, wherein the structural modeling software package includes FlexPepDock.

[0286] 33. The computer-implemented method according to any one of Examples 1 to 32, further comprising iteratively traversing a plurality of candidate peptides to identify at least one peptide having the maximum likelihood for forming the peptide-protein complex.

[0287] 34. The computer-implemented method according to any one of Examples 1 to 33, wherein the at least one peptide comprises at least a portion of a peptide expressed in normal cells of the subject.

[0288] 35. The computer-implemented method according to any one of Examples 1 to 34, wherein the at least one peptide comprises at least a portion of a neoantigen expressed from tumor cells of a cancer patient.

[0289] 36. The computer-implemented method according to any one of Examples 1 to 35, wherein the at least one peptide comprises at least a portion of a peptide expressed in a virus or bacteria.

[0290] 37. The computer-implemented method according to any one of Examples 1 to 36, wherein the at least one protein comprises a major histocompatibility complex (MHC) class I protein.

[0291] 38. The computer-implemented method according to any one of Examples 1 to 36, wherein the at least one protein comprises a major histocompatibility complex (MHC) class II protein.

[0292] 39. The computer-implemented method according to Example 37 or Example 38, further comprising iteratively traversing a plurality of candidate peptide sequences or portions thereof expressed in normal cells from the subject for at least one MHC class I or II protein, to order the peptide sequences or portions thereof according to the likelihood of formation of a peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complex.

[0293] 40. The computer-implemented method according to Example 37 or Example 38, further comprising iteratively traversing a plurality of candidate neoantigen sequences or portions thereof expressed in tumor cells from said cancer patient for at least one MHC class I or II protein, to order said neoantigen sequences or portions thereof according to the likelihood of formation of peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complexes.

[0294] 41. The computer-implemented method according to Example 37 or Example 38 further comprises iteratively traversing a plurality of candidate peptide sequences or portions thereof expressed in viruses or bacteria for at least one MHC class I or II protein, to order the peptide sequences or portions thereof according to the likelihood of formation of a peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complex.

[0295] 42. The computer-implemented method according to Example 40, further comprising developing a personalized anticancer vaccine based on the ordering of the neoantigen sequence or portions thereof.

[0296] 43. The computer-implemented method according to Example 39 or Example 41, further comprising developing an antiviral vaccine, an antibacterial vaccine or an autoreactive vaccine based on the ordered arrangement of the peptide sequence or portions thereof.

[0297] 44. A system comprising: One or more processors; and A memory communicatively coupled to the one or more processors and configured to store instructions that, when executed by the one or more processors, cause the system to perform a computer-implemented method according to any one of Embodiments 1 to 43.

[0298] 45. A non-transitory computer-readable storage medium storing one or more programs, the one or more programs including instructions that, when executed by one or more processors of a system, cause the system to perform a computer-implemented method according to any one of embodiments 1 to 43.

[0299] This description provides only preferred exemplary embodiments and is not intended to limit the scope, applicability, or configuration of this disclosure. Rather, the description of preferred exemplary embodiments will provide those skilled in the art with a feasible description for implementing various embodiments. It should be understood that various changes can be made to the function and arrangement of the elements without departing from the spirit and scope set forth in the appended claims.

Claims

1. A computer-implemented method for generating multiple compatible structures of peptide-protein complexes, the method comprising: The peptide sequence data of at least one peptide and the protein sequence data of at least one protein are input into a trained machine learning model to determine: The predicted pairwise distance data between at least one peptide amino acid residue of the at least one peptide and at least one protein amino acid residue of the at least one protein in the peptide-protein complex; and Predicted dihedral data for at least one peptide amino acid residue in at least one peptide of the peptide-protein complex; The initial structure of the peptide-protein complex is identified based on the predicted pairwise distance data and the predicted dihedral angle data. as well as The plurality of compatible structures of the peptide-protein complex are generated based on the initial structure, the predicted pairwise distance data, and the predicted dihedral angle data.

2. The computer-implemented method of claim 1, further comprising predicting at least one auxiliary feature of the peptide-protein complex based on the plurality of compatible structures of the peptide-protein complex.

3. The computer-implemented method of claim 1 or claim 2, wherein the predicted pairwise distance data includes one or more predicted posterior distributions of pairwise distances, and the predicted dihedral data includes one or more predicted posterior distributions of dihedral angles, and each compatible structure of the peptide-protein complex conforms to one or more posterior distributions of the predicted pairwise distance data and one or more posterior distributions of the predicted dihedral angle data output by the trained machine learning model.

4. The computer-implemented method according to any one of claims 1 to 3, wherein the trained machine learning model has been at least partially trained to predict at least one auxiliary feature of the input peptide-protein complex.

5. The computer-implemented method according to any one of claims 4, wherein the at least one auxiliary feature includes prediction of the binding of the at least one peptide to the at least one protein.

6. The computer-implemented method according to any one of claims 4 to 5, wherein the at least one auxiliary feature includes the prediction of at least one interface feature of the peptide-protein complex.

7. The computer-implemented method according to any one of claims 1 to 6, wherein the peptide sequence data includes a peptide embedding representation of the at least one peptide, the peptide embedding representation being generated using a trained peptide embedding machine learning model.

8. The computer-implemented method according to any one of claims 1 to 7, wherein the protein sequence data includes a protein embedding representation of the at least one protein, the protein embedding representation being generated using a trained protein embedding machine learning model.

9. The computer-implemented method of claim 8, wherein the trained peptide embedding machine learning model and the trained protein embedding machine learning model are the same model.

10. The computer-implemented method of claim 8 or claim 9, wherein the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model comprises: A base model; or a machine learning model that has been trained in a self-supervised manner using unlabeled protein and / or peptide sequence data to learn embeddings from input sequences.

11. The computer-implemented method according to any one of claims 8 to 10, wherein the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model comprises a deep learning model, a deep neural network, or a large language model (LLM).

12. The computer-implemented method according to any one of claims 8 to 11, wherein the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model comprises an evolutionary scale model (ESM).

13. The computer-implemented method according to any one of claims 8 to 12, wherein the trained peptide embedding machine learning model and / or the trained protein embedding machine learning model are integrated with the trained machine learning model and retrained simultaneously with the trained machine learning model.

14. The computer-implemented method according to any one of claims 1 to 13, wherein the trained machine learning model comprises a trained neural network.

15. The computer-implemented method of claim 14, wherein the trained machine learning model comprises a trained multilayer perceptron (MLP), a trained attention-based neural network, or a trained state-space model.

16. The method according to any one of claims 1 to 15, wherein the trained machine learning model is trained using a training dataset, the training dataset comprising incomplete structural data of one or more peptide-protein complexes.

17. The computer-implemented method according to any one of claims 1 to 16, wherein the trained machine learning model is configured to predict parameters characterizing: the distribution of pairwise distances between at least one peptide amino acid residue of the at least one peptide in the peptide-protein complex and at least one protein amino acid residue of the at least one protein; and the dihedral angle ψ of at least one peptide amino acid residue in the at least one peptide in the peptide-protein complex. The joint distribution of .

18. The computer-implemented method according to any one of claims 1 to 17, wherein the machine learning model is trained in the following manner: Receives peptide sequence data of multiple peptides; It can receive protein sequence data for a variety of proteins; Receives peptide-protein binding data for multiple peptide-protein combinations; A classifier is trained based on the peptide sequence data, the protein sequence data, and the peptide-protein binding data to predict peptide-protein binding for specific combinations of peptides and proteins. as well as The regressor head is attached to at least a portion of a trained classifier and trained based on pairwise distance data and dihedral angle data derived from crystal structure data of a set of exemplary peptide-protein complexes to predict the probability distribution of the pairwise distance data and dihedral angle data of the peptide-protein complexes.

19. The computer-implemented method of claim 18, wherein during training of the regressor head, pairwise distance data and dihedral angle data derived from crystal structure data for each combination of individual peptide amino acid residues and individual protein amino acid residues in an exemplary peptide-protein complex are presented as separate training instances.

20. The computer-implemented method of claim 18 or claim 19, wherein the regressor is further trained based on enhanced pairwise distance data and dihedral angle data.

21. The computer-implemented method of claim 20, wherein the enhanced pairwise distance data and dihedral angle data are data not yet experimentally determined, and / or data obtained by using a trained structure prediction machine learning model to obtain predicted crystal structures of peptide sequences and protein sequences known to bind together to form peptide-protein complexes, and / or data obtained in the following ways: Receive peptide sequence data and protein sequence data of at least one peptide and at least one protein known to bind and form a peptide-protein complex; The peptide sequence data and protein sequence data of the at least one peptide and the at least one protein are processed using a previous version of the trained machine learning model to determine the uncertainty of the pairwise distance data and the uncertainty of the dihedral angle data of the peptide-protein complex. The determined uncertainty of the pairwise distance data and the determined uncertainty of the dihedral angle data are applied to the structure of the exemplary peptide-protein complex to generate multiple compatible structures of the peptide-protein complex. as well as Pairwise distance data and dihedral angle data of multiple compatible peptide-protein complex structures were extracted to be used as augmented training data for the regressor.

22. The computer-implemented method according to any one of claims 1 to 21, wherein identifying the initial structure of the peptide-protein complex based on the predicted pairwise distance data and the predicted dihedral angle data comprises: Multiple crystal structures of peptide-protein complexes were obtained from protein structure databases; as well as The crystal structure is selected from the plurality of crystal structures based on sequence similarity, or based on the similarity between the pairwise distance data and dihedral angle data determined for the crystal structure and the predicted pairwise distance data and predicted dihedral angle data determined by the trained model.

23. The computer-implemented method of claim 22, wherein the protein structure database comprises a protein database (PDB), and / or wherein selecting experimental structures comprises: (i) For each exemplary peptide-protein experimental structure in a plurality of experimental structures, the distances and dihedral angles in the peptide-protein experimental structure are determined using the predicted distributions of pairwise distances and dihedral angles of the peptide-protein complexes output by the trained machine learning model; (ii) Select an exemplary peptide-protein experimental structure with the highest likelihood.

24. The computer-implemented method according to any one of claims 1 to 23, wherein generating the plurality of compatible structures of the peptide-protein complex based on the initial structure, the predicted pairwise distance data, and the predicted dihedral angle data comprises: Input the initial structure of the peptide-protein complex into a structure modeling software package; The predicted pairwise distance data and the predicted dihedral angle data determined by the trained machine learning model for the peptide-protein complex are input into the structure modeling software package; as well as The structure scoring function is iteratively sampled and evaluated from a plurality of possible structures of the peptide-protein complex to identify the plurality of compatible structures of the peptide-protein complex, wherein the compatible structures of the peptide-protein complex are structures that are (i) compatible with one or more constraints imposed by the predicted pairwise distance data and the predicted dihedral angle data, and (ii) optimize the structure scoring function.

25. The computer-implemented method of claim 24, wherein iterative sampling from the plurality of possible structures of the peptide-protein complex is performed using a Markov chain Monte Carlo (MCMC) sampling algorithm.

26. The computer-implemented method of claim 25, wherein the Markov chain Monte Carlo (MCMC) sampling algorithm includes the Metropolis-Hastington algorithm.

27. The computer-implemented method of claim 25, wherein the Markov Chain Monte Carlo (MCMC) sampling algorithm includes the Hamiltonian Monte Carlo algorithm.

28. The computer-implemented method of claim 25, wherein the Markov Chain Monte Carlo (MCMC) sampling algorithm includes the Langevin Monte Carlo algorithm.

29. The computer-implemented method according to any one of claims 24 to 28, wherein the structural scoring function includes a proxy for free energy calculation.

30. The computer-implemented method according to any one of claims 24 to 29, wherein the structure scoring function comprises a proxy for calculating the free energy of peptide-protein binding, a proxy for calculating the free energy of at least one peptide-protein interface feature, or any combination thereof.

31. The computer-implemented method according to any one of claims 24 to 30, wherein the scoring function is based on one or more of the following: Let's start from the beginning. Quantum mechanical calculations, density functional theory (DFT) calculations, semi-empirical calculations, molecular mechanical force field calculations, statistical potential calculations, neural potential calculations, or machine learning models trained on structural data.

32. The computer-implemented method according to any one of claims 24 to 31, wherein the structural modeling software package includes FlexPepDock, and / or wherein the one or more constraints imposed by the predicted pairwise distance data and the predicted dihedral angle data are included in the structural scoring function.

33. The computer-implemented method according to any one of claims 1 to 32, further comprising iteratively traversing a plurality of candidate peptides to identify at least one peptide having the maximum likelihood for forming the peptide-protein complex.

34. The computer-implemented method according to any one of claims 1 to 33, wherein the at least one peptide comprises at least a portion of a peptide expressed in normal cells of the subject.

35. The computer-implemented method according to any one of claims 1 to 34, wherein the at least one peptide comprises at least a portion of a neoantigen expressed from tumor cells of a cancer patient.

36. The computer-implemented method according to any one of claims 1 to 35, wherein the at least one peptide comprises at least a portion of a peptide expressed in a virus or bacteria.

37. The computer-implemented method according to any one of claims 1 to 36, wherein the at least one protein comprises a major histocompatibility complex (MHC) class I protein.

38. The computer-implemented method according to any one of claims 1 to 36, wherein the at least one protein comprises a major histocompatibility complex (MHC) class II protein.

39. The computer-implemented method of claim 37 or claim 38, further comprising, for at least one MHC class I or II protein, iteratively traversing a plurality of candidate peptide sequences or portions thereof expressed in normal cells from the subject, to order the peptide sequences or portions thereof according to the likelihood of formation of a peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complex.

40. The computer-implemented method of claim 37 or claim 38, further comprising iteratively traversing a plurality of candidate neoantigen sequences or portions thereof expressed in tumor cells from said cancer patient for at least one MHC class I or II protein, to order said neoantigen sequences or portions thereof according to the likelihood of formation of peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complexes.

41. The computer-implemented method of claim 37 or claim 38, further comprising iteratively traversing a plurality of candidate peptide sequences or portions thereof expressed in viruses or bacteria for at least one MHC class I or II protein, to order the peptide sequences or portions thereof according to the likelihood of formation of a peptide-MHC class I (pMHC-I) or peptide-MHC class II (pMHC-II) complex.

42. The computer-implemented method of claim 40, further comprising developing a personalized anticancer vaccine based on the ordering of the neoantigen sequence or portions thereof.

43. The computer-implemented method of claim 39 or claim 41, further comprising developing an antiviral vaccine, an antibacterial vaccine, or an autoreactive vaccine based on the ordered arrangement of said peptide sequences or portions thereof.

44. A system comprising: One or more processors; and A memory communicatively coupled to the one or more processors and configured to store instructions that, when executed by the one or more processors, cause the system to perform a computer-implemented method according to any one of claims 1 to 43.

45. A non-transitory computer-readable storage medium storing one or more programs, said one or more programs comprising instructions that, when executed by one or more processors of the system, cause the system to perform a computer-implemented method according to any one of claims 1 to 43.

Images